I was sitting in a coffee shop earlier today, when the waitress asked where I was from. A fella over at the next table caught the accent, and we got talking. Somewhere in the next few minutes he mentioned he’d worked in an AI lab during the AI winter of the late 80s, and I thought, since this random conversation seems to be happening anyway, let’s capture it. I asked if he’d be comfortable making it an interview format with transcription, so we agreed to sit down properly.

Robert Hooker has been working in IT since 1989, starting at Northwestern University’s Institute for the Learning Sciences and ending up in London, some thirty-odd years later, running the internal AI adoption programme at a social care charity. He says he’s watched this happen before.

So tell us a wee bit about yourself and how you got started.

My name is Robert Hooker. I’m a senior person in cloud computing. I have been working in the field of IT since late 1989. I first worked in the United States at Northwestern University’s Institute for the Learning Sciences. This was during the AI winter, so the Institute was actually an AI lab.

The researcher who ran it was Roger Schank, a famous early AI researcher, who I gather sadly features in the Epstein files. He was one of the people who introduced the concept of “scripts” in early AI. An interesting field, in a relatively failed branch of AI back then.

We were still in what I call the Minsky winter, when Marvin Minsky thought he had proven that neural nets would never work. So people weren’t working on them, and they were trying to teach computers how to think in Lisp and getting nowhere.

Just to check, the Minsky proof, that was the XOR problem?

Yeah, of course.

And so I was being trained that machine learning was just really stupid, but a lot of our people were very interested in it. It was a nightmare.

I do have a published academic paper, from a conference, Hooker and Slator, on a model of AI decision-making. It didn’t say AI, of course, because we were in the AI winter. But it was about how agents make decisions in an economic role, by the agents simulating pursuits.

And that was where I introduced Brian Slator, who went on to North Dakota State University to do the rest of his career research there.

But I didn’t stick around in academia. In the mid-90s I went to the private sector, going from web design, to SharePoint, to cloud, to Microsoft 365, and now Copilot. The guiding principle, which I’d recommend to all the young people in IT: if you can’t beat them, join them.

"For us older guys, this wasn't supposed to have worked. And it just blew our minds."

Have you come across the idea of Sutton’s Bitter Lesson? I work in the neural network space, and the idea addresses the frustrating trend that you throw more and more compute at the problem and it just somehow works better.

Oh yeah, that principle. Which wasn’t supposed to work. That’s what was amazing.

I gave an interview about 16 years ago on a podcast called The Future And You, and he asked me about predictions. And I did not say that. Some voice recognition, maybe, but the way I hedged it was: “I’ve been surprised before.”

When you work with a small neural net, there’s an optimal number of layers, there’s an optimal amount of training. But when you have a massive set of these weighted neurons, for us older guys, this wasn’t supposed to have worked. And it just blew our minds.

It took a while convincing. It took me several months of playing with OpenAI to be convinced that it actually was working. And here’s the thing. If you go to it determined to show it doesn’t work, it’s very easy. It’s very, very compliant in not working for you.

I suppose you get the answer you went looking for.

An interlude in Edinburgh

The reason we got talking in the first place was you’d mentioned being in Edinburgh for the 2014 referendum. What was that like, being in the buzz of it?

Oh, it was a lot of fun. My wife and I, my late wife, if I can make a plug for her, her name was Gail Orenstein. Just Google “female drone journalist”, she shows up. She died five years ago, which is unfortunate, but she was a very well-respected journalist. It was always a lot of fun with her.

I was working remote at the time, so I could go and spend a whole week in Edinburgh while working. And to be honest, this wasn’t like the Brexit referendum. It wasn’t ugly, it wasn’t bitter, at least in Edinburgh. There were some fisticuffs that we saw, and some crazy things being yelled back and forth, but for the most part I saw it as an amazing celebration of Scottish identity, tempered with Scottish practicality.

Ha, mind you, the fisticuffs might just have been a typical night out in Edinburgh.

Or even lighter than a typical night out in Edinburgh.

I also do art, and I found it artistically very inspiring. Edinburgh is just inspiring to begin with, but then to see all this going on as well. I was doing illustrations where I brought Edinburgh alive by merging the statues with the people. My drawings included the birds on their heads taking shifts, because that’s what you actually see.

And one of the great joys of being a journalist is that you’re an observer. The thing that really struck me, comparing it to the Brexit nightmare, was a wedding photograph where the bride was Yes and the father was No Thanks. It was a great photograph. Brexit was never like that. Brexit was just ugly.

Aye, definitely a more alienating experience, at least from how I saw it.

The dot-com bubble, and who held their nerve

You mentioned being around for the dot-com bubble. What was that like?

In the late 90s it was great. A lot of my friends were working as contractors in IT, and deployment was still pretty limited and very development-focused. Two things got the IT community involved.

One, it was the first time something was being delivered where you didn’t have to install anything on the end machines. The same bit of work could go on an Apple. People forget how revolutionary that is.

Secondly, it was the introduction of lightweight languages that weren’t C or C++. You could run websites on PHP, which was a bit poor, but it was very revolutionary at the time. Then ASP came out, and that took off too. Java Server Pages were too heavy-handed and didn’t take off. And then JavaScript, and you could separate what was happening on the client.

So suddenly there was an easy level of entry if you were a technical person. And suddenly there was demand for graphic art skills, which there hadn’t been before. All of our friends in the late 90s were suddenly making a lot of money. That’s when I moved to England with my wife, because she was a photojournalist and wanted to be in England for obvious reasons.

But the investment was way over the line. And when that dried up, for a few years you couldn’t get any work. It was really, really hard. The bomb hit in 2000, and by late 2002 it was as much because of 9/11 as it was because of the internet. The financial system just stopped funding. But within three years it was back up and running.

So what’s the lesson you’d draw for what could be called the contemporary AI bubble?

The people who didn’t panic and stuck with the internet during the dot-com bomb, when all the brains were saying it was bad, the Bezoses and the Zuckerbergs, they’re the richest people on the planet right now. So better to be closer to the people who held their nerve, if you have a choice. I’ve been there before.

There’s probably a good few AI Pets.com knocking about right now. But Amazon came out the other side looking not bad at all.

Maybe. But there is a difference. Back in the day you could create a website on a single server plugged into the internet. You could pay under $1,000 for the server, run e-commerce on it, deliver solutions on a pressed CD. It’s not like that now. You’ve got big players who are really well-heeled. Amazon, Microsoft, Google, and Meta all have revenue streams. Those are the big players in AI.

I famously missed the social media wave, by the way. I had a nephew who attended Harvard at that time, and he comes in and says “a friend of mine is creating a social network”, and I said, “Microsoft and Yahoo are dominant.”

Auch, geezo. That’s a story to tell at least.

What he’s doing now

So what is it you’re doing now?

Right now, my wife passed away five years ago, so I kind of slowed down. But I’m working for a not-for-profit called Dimensions, which just won an award for being the best workplace in social care and not-for-profit in the UK. We’re the UK’s biggest not-for-profit provider of support for adults with learning disabilities and autistic people, but we don’t even have 3% of the market. So even though we’re the biggest, and we grow like crazy, we’re not quite sure where to go in the future.

I was hired about four or five years ago to work with cloud computing and architecture. But then AI got introduced, and so what I’ve been doing for most of the last two years is overseeing our internal AI adoption project.

And how’s that going?

Very interesting, but also very challenging, especially with security teams. They want some kind of framework to make sure their AI is compliant. And I keep telling them: these frameworks don’t work as of yet.

I was at a conference by one of the vendors, and I kept asking, “look, I’m building these AI agents. I’ve been told by security that they’re concerned, without them telling me exactly why. So how do I structure my project?” And all they say is “buy this product.” No. That’s not going to fix it.

I know from being an architect, you have your business requirements statement, and a process that takes you from there to something you can test. That isn’t here in AI as of yet. The big problem with an agent business requirement is that people don’t know to ask for a chatbot that has cognitive skills in a certain area. So we get people to understand LLMs, then get them to understand what agents can do. And then suddenly the use cases come out of the roof.

What are the concerns, especially given you’re working with quite sensitive data?

First of all, I’m doing it very differently from a lot of the big businesses. I’ve talked to friends of mine who are concerned, in large part because the companies are coming in and deploying it as if it’s an old factory where you put in new machinery, kick the staff out and put security guards in front of it. That isn’t how we’re doing it, and I don’t think that’s going to work.

Because we’re a not-for-profit, almost everybody working with us could make more money somewhere else and they choose to stay anyway. So we have a motivation factor that isn’t just the money. That gives us an opportunity, because they actually trust us.

So we’re getting people who are overwhelmed, constantly having to file this report and that report, keeping the news updates going, updating the web page and the intranet page, underpaid and we can’t hire new people. And I’m saying: alright, here’s Copilot, this is what it can do, ask me about it. And then we get them to discover the uses.

So it’s grassroots? You’re trusting your colleagues’ expertise and seeing where you can cut, so they can focus on what actually matters?

Yeah. They’re focusing, and I’m out there to support them with the technology.

And the compliance side of that?

We’re constantly getting more and more regulations about how to work. We have to have a person whose only job is to tell us what’s GDPR compliant and what isn’t. So we have to spend money and effort on it, and keep going on it. It’s kind of a pain, but we do it.

Mind you, the punishment for not bothering is really just your reputation. If you’re determined to keep yours, you do it properly. So it ends up being a kind of self-imposed punishment for the people who decide to do it the right way.

And the thing is, you can also use GDPR to annoy people. You just go on an AI and it can write a GDPR request, and you can inundate somebody you’re mad at with them. Although I’ve spoken to the people who handle ours, and they don’t think they’re getting many of those. Ours are mostly sincere, very often because somebody’s moved house and they just want to make sure their record is complete.

So the future of agents is folk flooding their competitors with GDPR requests using agents, and then those agents responding to all of it.

They’ve been doing that with lawsuits forever.

Aye, lawsuits, patent trolling. It’s cheaper to annoy.

You don’t have to be a rich person to make somebody’s life miserable.

Why the code monkey is going away

"You have to become an expert in every agent you write."

Has building agents changed how you’d approach a project?

One of the things you learn very quickly with agents is that the old traditional “give me the requirements, I’ll write the use cases, and I’ll code it” does not work in cognitive computing. You have to become an expert in every agent you write. If it’s a finance agent, you have to become an expert in finance.

There was a project manager who wanted an agent, and at first I thought I’d just point it at all their templates and processes. And it was okay. But to get it to work appropriately, what I had to do was write proper ontologies into the RAG to explain how the work actually worked, and run it through, and run it through.

I end up having to learn, which I enjoy, but I think in the future a good programmer in AI isn’t going to be a code monkey. Code monkey is going to be a career that goes down. It’s going to be somebody who works in a certain field and is really good at cognitive descriptions of that field. Somebody who can understand what cognitive skills are necessary to do a particular job.

Which pushes AI more into psychology, sociology, and literature. I have an undergraduate degree in cognitive science, and that’s been much more useful than anything else in the current work. Particularly behavioural theory.

Go on, what carries over?

One rule of cognitive theory is that it’s not enough to say no. If you’re correcting a behaviour that’s wrong, it’s very hard to just tell someone to stop doing it. You have to tell them the proper behaviour instead.

So when you do a RAG design and your instructions are just “don’t do this, don’t do that”, it’s not going to be as effective as one that says “do this rather than that.” That’s solid psychology.

Computer programmers think if-then, if-then. And it didn’t matter if it was negative or not. Nobody worries about that. But human beings don’t work that way. What works for humans is “not this, rather this.”

Aye, rather than the strict logic of C, you’re almost shaping a surface for it to slide towards what you want.

I don’t know what’s going on in the industry side of it, but I do know from behavioural cognitive studies that this has been established for almost two centuries now as the right way to go.

That resonates. Two years ago 80% of my job was writing code; now maybe it’s 5%. A lot of it is managing agents, and another part is just talking to people.

Mine was the same. About five years ago 80% of my job was writing documents that nobody would read. Now I do none of that. Now I’m talking to people.

Hallucination, or imagination

"In probably 10 years they'll rename hallucination to 'imagination'. Because that's really what it is."

What about the risks?

The back-end ones we’ve covered. The big risk on the front end is, of course, hallucination. And you can lecture all you want, but people are going to have to learn from their own negative experience. Unfortunately, if you go on social media, people think that one negative experience proves the whole thing doesn’t work.

But this is one of the things I’ve learned. I’ve had some agents where I put in such guardrails that they wouldn’t hallucinate. The problem was that they couldn’t generate new thinking either.

Right, you make it so safe it can’t say anything useful.

I just thought hallucination was bad. But look, you’re a human being. I may ask you a question about something you’ve never dealt with before, and you may imagine some solution. You’re going to have to test it, but you might come up with something.

So you need to look at the domain. Is this something that needs a precise legal answer? Turn hallucination down. Other things, like drafting job adverts or public announcements, those things need to hallucinate some.

In probably 10 years they’ll rename hallucination to “imagination”. Because that’s really what it is. And people will just take it for granted: don’t let the thing’s imagination get away with it. You’ve got to contain its imagination. I think it’s a better term.

Any other experiences that have made a difference?

The biggest one is getting people to not write prompts, but to write prompts that write prompts. Write a prompt to write a prompt, then review that, then run the prompt. That’s the best way to go. The AI speaks AI better than you.

Cheers Robert, one of the more fruitful random coffee shop conversations I’ve had.

He had a sketchbook with him, and got me to photograph a couple of pages before I left. Not the Edinburgh statues he’d described (from over a decade ago), but pen studies of insects, flowers, shells, and shapes coming apart.

We chatted a bit more off the record after that, and then I finished my coffee and carried on with my day.

A few years ago, I started treating it like a systems problem and built a scheduling layer around it. Motivated by a recent chat with some interested parties, this post briefly documents that system, how it broke, and where I’ve landed.

Some philosophy

I dinnae want to turn this post into a pontification on humans as social creatures, and compare and contrast (possibly outmoded) models of evolutionary psychology vs contemporary and future-facing aesthetics of the woven nets of relationships that we want to build. Buy me a coffee and yap on that one-on-one.

However, keeping up with folk is something I like to do, and I believe in the power of my own agency: that I can decide the things I want to do, and have some understanding of why I want to do it.

Once I’ve exercised the agency of choice, I’m keen to minimise the number of low-agency steps required to achieve that. This is why I like having keyboard shortcuts for common tasks: intent, action, and result should all be closely linked.

Given that, exercising the agency of “I want to maintain this relationship with someone that I’m not constantly thinking about” (e.g., not a partner or immediate family member) can be thought of as a systems problem, intersecting with the social dynamics of it all. Understanding what stage a relationship is at, if you or the other person just isn’t feeling it and you should back off, and what you should be talking about is all something I want to be using my own judgement for.

Or put simply, I want to be the one actually living my relationships, any tooling around it should just make that easier.

We can automate the other parts, but that seems a bit grim. Born, push a button to do all the things in your life, die.

In typical tech person fashion, let’s quickly make the evil thing, because it can be done. I beg you not to use the following script:

import anthropic
from people import Person, is_due


def build_prompt(person: Person) -> str:
    history_block = "\n".join(f"- {m}" for m in person.message_history[-3:]) or "None"
    return f"""
You are helping someone maintain their relationships. Draft a short, warm, natural message to send to:

Name: {person.name}
Relationship: {person.relationship}
Contact frequency goal: {person.frequency}
Last contacted: {person.last_contacted}
Recent message history:
{history_block}

Keep it brief (2–4 sentences). Don't be generic or sycophantic. Match the tone to the relationship.
Make it genuine, witty, and not weird.  The feeling of living in an alienated dystopia where even the
connections between friends has melted away to tokenisation should *not* come across.
""".strip()


def generate_message(client: anthropic.Anthropic, person: Person) -> str:
    response = client.messages.create(
        model="claude-opus-4-5",
        max_tokens=256,
        messages=[{"role": "user", "content": build_prompt(person)}],
    )
    return response.content[0].text.strip()


def run(people: list[Person]) -> None:
    client = anthropic.Anthropic()

    for person in people:
        if not is_due(person):
            print(f"⏭ Skipping {person.name} — not due yet\n")
            continue

        print(f"✉ Generating message for {person.name} ({person.relationship})")
        msg = generate_message(client, person)
        print(f" → {msg}\n")

The original system: Emacs org-mode

My first serious attempt was back in late undergrad, and used org-mode in Emacs.

Emacs is a family of text editors that traces its roots back to the 1970s, originally developed at MIT. The GNU Emacs branch, maintained by Richard Stallman — of the Free Software Foundation and GNU Project fame — has been the dominant strain since the 1980s and remains actively developed today. It’s less a text editor than a Lisp runtime that happens to be very good at editing text.

Stallman at FOSDEM 2005. Photo by chrys, CC BY 2.0.

Org-mode has a built-in scheduling syntax that lets you attach dates and recurrence patterns to any heading. The agenda view then collects everything due today across all your files — a kind of personal dashboard.

A contact entry might look like this:

* Vangelis Ogunkunle
  SCHEDULED: <2026-01-15 Thu +4w>
  :PROPERTIES:
  :CATEGORY: folk
  :END:

The +4w suffix is a repeat interval: once you mark the task done, it reschedules itself four weeks forward. You can also use ++ (shift by interval from today, not from the original date) or .+ (shift from today, but only forward) depending on how strict you want to be about drift:

SCHEDULED: <2026-01-15 Thu ++4w>   ; always lands on a fixed cadence
SCHEDULED: <2026-01-15 Thu .+4w>   ; 4 weeks from whenever you actually did it, my preferred approach for this

Running my shortcut for M-x org-agenda and pressing a gave me a daily/weekly view of everyone I was due to speak to. It was great. The system lived in a single folk.org file, was version-controlled with the rest of my config, and required zero internet access.

I would take a look at my agenda, see who I felt like messaging that day, and adjust the schedule frequency if I felt it needed to be made more or less frequent.

A typical day might look something like:

Sunday     12 April 2024
================================================================================
 Folk
  folk:                      Lorna Rauf
  folk:                      Aunt Redacted
  folk:                      Akshil Wierzbicki

 General Tasks
  agenda:                    Watch Welder talk
  agenda:                    libclang exploration
  agenda:                    ALIFE conference review

Crucially, I’m not obliged to message a person on that day, they will roll over until I actually check them off.

The agenda view is just one piece of it — org-mode and Emacs more broadly are very configurable through their built-in customisation system and init file. Want custom agenda commands that filter by tag, show only people you haven’t contacted in over a month, or group by city? All doable. The price of admission is a willingness to spend time in the parenthesis mines of Lisp.

How it fell apart

I moved to London in 2024, and suddenly found myself with a tube commute, and a corpo laptop. I didn’t feel comfortable putting my social web on something that could ostensibly come out in discovery or the whims of company IT.

I also found I wasn’t using my personal laptop daily, I was out and about doing stuff in my off-hours. As a result, I found a few months in “hey, I’ve not been maintaining some of the relationships that I care about, why is that?”.

I realised that my system was reliant on daily use of my personal laptop, and the time that I was free to message and catch up with people was on an underground train, on mobile, with minimal internet connection.

Alas, Emacs as I use it is not well suited to the mobile experience, so I needed a stop-gap solution.

Google Calendar Tasks: a workable compromise

Google Tasks (built into Google Calendar, and also as a standalone app) turned out to be a reasonable middle ground. It syncs to my phone, works offline on the underground, and supports due dates with recurrence. I’ve given social tasks their own colour, to make them distinct from other tasks and calendar events.

The workflow is simple: each person gets a task, set to repeat on whatever cadence makes sense. When I’m on the tube, I can take a look at my Google calendar for the day, and see who’s up.

Day view — folk tasks show up in yellow alongside regular calendar events.

Adding a new contact task is just a tap on the + button, then selecting “Task”:

The new item menu — Task sits alongside Event and other options.

Each task gets a repeat interval, set at creation:

Setting a 6-week repeat on a contact task.

It doesn’t have org-mode’s expressive repeat syntax, and the wide scope of additional configuration I’d like to have. But it’s good enough, and critically, it’s always with me.

The future: getting back to org-mode

The org-mode approach is still the one I’d prefer. The plain-text format, the expressive scheduling syntax, increased privacy, the dork creds, and the composability with the rest of my Emacs workflow are hard to replicate. The question is just whether I can get it onto my phone reliably enough to replace Google Tasks for the commute use case.

In an era of agentic coding, sure it will be way easier, however the question is how arsed I can be, given this approach seems to be “good enough” for how my life is going right now.

The legible artefacts of notes and software tools accrue in their home directories, most of which has not been fully read by their eyes. But they trust that it sufficiently represents their will, and it may have its use down the line.

They may pass the audio of their meetings or conversations through models, for summarisation or analysis. The motivated will have live interaction, consulting their models for potential insights or questions to ask. With practice, the pauses and darting eyes are difficult to spot.

The personal agency of the token cyborg will vary. Some wear their models like exo-skeletons, strengthening their own resolve, and amplifying their will. Others are worn by the models, productive half hours passing with nary a conscious thought.

The token cyborg hungers for larger context windows, longer reasoning chains, higher throughput, lower latency. In an hour they may crunch through more floating-point ops than the whole human race did until around the time of the Chernobyl disaster.

Self-hosted, or heavily subsidised TaaS, the token cyborg shifts their infrastructure often. How much of themselves exists in a data lake to be picked over? As their trust in their tools grow, they expose more of themselves. An email API key here, a family group chat there. Context is power.

Desires and goals now form as prompts, quick references for the context to fetch — trust that the intent is evident. An implicit knowledge of models’ pitfalls, caveats and requirements sprinkled in at the end of requests. Make it efficient, don’t make mistakes, don’t hallucinate. Make it sound charming.

A WiFi-less flight is to have one’s arms bound.

Dr Dr Mahnoz Illias completed her PhD at the University of Glasgow in 2025. Aye, she really is a double doctor 😎, originally earning a medical degree in Bangladesh, before completing an MSc in Global Mental Health at the University of Glasgow, supported by the Commonwealth Scholarship Commission. Her PhD research examined policing as a gendered organisation and the impact on the health and well-being of female police staff and officers in the UK. Here, I had a wee chat with her about her research and perspectives.

Awright Dr Dr Mahnoz?

Hello there Dr Perry Josef Gee Gibson.

What was your research topic about?

I looked at the gendered toll of policing on the health and well-being of female police officers in the UK.

Who were you working with for that?

I was working with my supervisors and the wider team that I engaged with. My primary supervisor was based at the School of Health and Wellbeing at the University of Glasgow, and my second supervisor was at the Adam Smith Business School.

It was quite an interdisciplinary project. I was based within these two schools, but a lot of the work I did was actually closer to the field of criminology rather than either public health or business/organisation studies.

Outside of the school, I worked closely with the Scottish Women’s Development Forum of Police Scotland. They’re a group who represent women within Police Scotland — all the women staff and officers. It’s a voluntary organisation because, although police forces keep saying they’re “trying their best to accommodate the welfare of their employees”, but claim they “do not have the funds” for formal support. So volunteers step forward to fill the gap. I also worked with broader groups like the Scottish Institute for Policing Research and the Academic Team of Police Scotland.

What was the connection with the Adam Smith Business School; I don’t immediately see the connection to the research area?

It was mainly about the organisation — the culture, structure, and norms, and how they impact health and well-being. Although in public health we do study this (and public health itself is very interdisciplinary), it’s mainly the people in “organisation studies” who study how organisations are functioning. They’re located at the business school, so that’s where the Adam Smith Business School comes into my research journey.

"Social support, especially workplace social support, plays a huge role in our lives, which we often overlook."

You were originally a medical doctor, then you did a Master’s in Global Mental Health. How was that perspective shift—looking at mental health not just as an individual matter, but as a large-scale policy question?

It was different. When we talk about health or mental health, what I’ve found is that people often think it has to be related to institutional medicine, or hospitals, or doctors dealing with it. But in our day-to-day life, we can see how each individual has a health experience that is much less formal. Our eating choices, our social interactions, our work environment—all of these impact our health.

We also share a lot of experience from the people around us, which impacts us both mentally and physically. Acknowledging this in the workplace is important — in the past, these issues were often presented as individual problems. But I believe we’re coming to recognise it as a shared responsibility. Every interaction—including this conversation—happens in a shared space. Both of us bring our accumulated knowledge and experiences to it.

I also feel that when we talk about health, we normally try to narrow it down to medical reports or such things, but there is something way beyond that. Health is not just a medical issue—it’s a social issue, it’s a justice issue. These things came up again and again in my research.

Taking the perspective that not everything can and should be medicalised?

Yes, when I finished medical school in Bangladesh, my idea about mental health was quite narrow. When I talked about mental health, I used to think it had to be something major—schizophrenia, bipolar disorder, or something that needs regular medication and to be dealt with on a regular basis.

Outside of that, our understanding from our education was that “depression is there”, and people just have to deal with it until they need medication. What I’ve come to understand is that mental health isn’t separate from physical health—they’re intertwined, happening at the same time. It’s not a separate entity; it’s part of who we are. That means we should be aware of it in our everyday lives, rather than just in crisis.

Was your research focused on Scotland, or did it explore the wider UK?

It was actually throughout the UK. However, as the project received half of its funding from the Scottish Institute for Policing Research—and I’m grateful to them and MRC UK for their support—I had more connection with organisations based in Scotland. But I did a mixed-method study. I had to analyse two large national cohorts of police officers, which covered information from all police officers and staff across the UK—it wasn’t confined to Scotland, and I did not confine my results to Scotland either.

When I did my primary data collection, most of the participants were from Scotland—probably two-thirds—and the rest were from across the UK.

In brief, what were the main findings of your thesis?

I would say the main findings were that the health and well-being of female police officers cannot be understood as a single piece of work. Rather, I would describe it as a puzzle. By doing research on all these different parts, I got a closer picture of the entire puzzle, but it’s still not solved. It was more like a mosaic, with pieces fitting into each other rather than isolated factors—the social aspect of health working one way, the mental health aspect another way, the physical aspect there. They’re all intertwined.

All of these health and wellbeing issues—stemming from the cultural and structural issues—often, as I said earlier, we try to describe them as medical problems. But health is not just a medical problem. It is a justice issue as well. What facilities you’re getting within your organisation, what you’re given or not given, can be seen from a justice point of view. At the root of everything, there is actually organisational injustice, which is controlling how it impacts overall health.

Lastly, I found that social support, especially workplace social support, plays a huge role in our lives, which we often overlook.

These seem like interconnected problems. How much can be attributed to being a woman, versus just “being a polis¹ is stressful”?

Every organisation or institution within a society, including families, is a reflection of the existing society itself. When I was talking to those police officers, one of my inclusion criteria was they would need to have at least ten years of experience working as a police officer—to reflect back on how the culture changed over time, how policies changed, how that impacted them.

All of them came into policing knowing there would be a certain level of operational stress. They knew what policing was—definitely there were some surprises like every other job, but gradually in day-to-day life, the relational aspects became more prominent.

There’s a scholar with a famous theory called “Gendered Organisations”—her name is Joan Acker. She explains it very nicely. She says that a “job” itself is an abstract idea; for its embodiment, you need a person to be there.

And that person, when they enter the role, they bring all of their gendered ideas about women with them—what they learned throughout their life about the limitations of women, what they can do, what they cannot do. They came with their perceptions and kept spreading those ideas. As society has changed, those ideas have also changed. But still, though we keep saying that men and women are equal in our society, that’s not quite the case. Policing remains a gendered organisation, and this has real impacts on the health and well-being of female police staff and officers.

"Health is not just a medical issue—it's a social issue, it's a justice issue."

Is it your impression that police in the UK have more of a “boys’ club” problem than other places? Or is it more in-line with the issues in wider society?

They definitely had their own boys’ club issues. They had cultures where women wouldn’t be invited to the informal parties or after-work gatherings, or they’d always assume that if you have children, you must need childcare.

Are there any articles folk could check out if they were interested?

I have two articles out so far, and two others are under review.

My meta-ethnographic review, which looked at the interplay between organisational injustice and the health and well-being of female police officers and staff across the globe, was published in BMC Public Health [1].

Another piece of work compared the impact of workplace social support on the mental health of parents, both men and women, published in the journal “Policing” [2].

[1] M. Illias, K. Riach, and E. Demou, ‘Understanding the interplay between organisational injustice and the health and wellbeing of female police officers: a meta-ethnography’, BMC Public Health, vol. 24, Sep. 2024, doi: 10.1186/s12889-024-20152-1.

[2] M. Illias et al., ‘Exploring gender differences in policing: the role of workplace social support on the mental health and wellbeing of parents in policing’, Policing, vol. 19, Jan. 2025, doi: 10.1093/police/paaf037.

Were there any technical tools in your work that you were using, and any frustrations with them?

Definitely. One of the datasets I was using was the “Airwave Health Monitoring Study”. Since it was quite sensitive, to access that data, I had to get access to a “safe haven”, or what they formally call a Trusted Research Environment. The operators had maintenance windows (thus, I couldn’t always access it when I needed to), and sometimes because it’s a big dataset hosted somewhere else, there were technical issues. Many days I was frustrated because you’re in the mood to analyse and then you cannot access the data.

When I began my PhD, I didn’t know how to code in R (a programming language for statistical analysis), but I needed to for its flexibility and the amount of data analysis I had to perform. Learning that was also a challenge. I used to freeze at the thought of having to engage in coding work—the word “coding” was very frightening. But after my analysis of the first database, I got quite confident, and it didn’t take me much time to do my second analysis.

Do you think you would have benefited from having access to more technical support?

I had people in my unit who I could reach out to, but in a PhD you can’t control which part of work you’re going to do at what time of day. If it was a friend, I could reach out anytime. But if it’s a colleague, it’s a formal relationship. As a new person within the unit, I was also hesitant to reach out to them. Whenever I had issues, I could reach out to people, but these were colleagues who had volunteered to help me—there weren’t any designated people you could always go to. Having someone like that would have definitely benefited me or people like me who have just started coding.

What was your experience navigating the research space in the UK as a Bangladeshi woman?

I didn’t face any issues particularly from a gender perspective. But compared to my MSc, I felt a bit isolated. I don’t know whether that was the nature of the PhD, or something was wrong with me, or I wasn’t ready—it was just after COVID as well, so people were quite isolated at that time. Maybe all those factors worked together.

I was still hanging out with my old friends outside that space, but compared to that, I did not have that many close friends within my own department. They were very lovely people, very cooperative, but somehow I didn’t have that connection with them.

But there was another space—the Scottish Institute for Policing Research—where I met a lot of people from different fields: criminology, cybersecurity, forensic backgrounds. I made good friends with them. I don’t know what the reason was behind this, but that’s how it was.

"Coming from Bangladesh, I expected the UK to be a 'utopia'... But when I looked beyond the surface, it was very shocking."

For any organisation in general—you’ve looked a lot into organisational structures. Is there any general advice you’d give to leaders or individuals about how to help manage the issues you’ve seen in your work?

One of the things was that Police Scotland is a single force, but in England they have around 43 forces. Across different stations or forces, they each were tackling different issues well, and others less well. But those were ad hoc initiatives, with success stories not being shared with each other.

I know there’s decentralisation of policing and those ideas in criminology, and there are certain reasons for that. But I felt they should have a common platform where they can share positive initiatives that are actually helping people deal with the issues they face in their day-to-day life.

If they can find a way to share those and adopt the effective programmes that are working for people, that would be really beneficial for the entire organisation or the entire police force across the UK.

You mentioned the PhD ended up being more about criminology. Could you briefly explain what you meant by that?

I would say it’s criminology because every time I was going to conferences and knowledge exchange programmes, I was mostly interacting with people from criminology or forensic psychology backgrounds. They’re the ones coming to work on what’s happening with policing, how policing can be better, rather than someone from a health background.

In that sense, I didn’t have much in common with other people in terms of background. But they were very welcoming, so it wasn’t an issue.

Also, a major portion of writing up my final thesis involved understanding the policing structure within the UK—what policies are out there, under what law police organisations in the UK came into being. Those are things more from criminology rather than my health perspective. They were new ideas for me.

Anything else you want to highlight?

Another interesting finding: when I started my project, I was initially expecting the focus to be about physical health. Many police officers are exposed to more danger than a typical office job, and if we’re talking about women—our physiology is generally different from men, and this has some impact on our musculoskeletal health. Perhaps my research would take me into questions around musculoskeletal problems they’re facing, and how the workplace could change to support that.

However, when I started to look into it, what came up more consistently was mental health issues. I was very surprised that no matter which part of the world the articles I was researching were coming from, they had similar issues: regardless of the culture—whether it was from the Global South or Global North. It was surprising for me because, coming from Bangladesh, I expected the UK to be a “utopia” where everything was perfect, or at least significantly better, women would be getting their rights properly, they didn’t have to fight for it, they were living their lives very comfortably.

But when I looked beyond the surface, it was very shocking and frustrating for me to deal with those accounts on a daily basis, and that these issues exist in the UK as much as they do in other parts of the world. It also helped me realise how seemingly gender-neutral policies are not as neutral as they seem.

Thank you Dr Dr Mahnoz

Thank you Dr Perry Josef Gee Gibson.

I was delighted to be invited to both give a talk, and participate in the panel discussion at the CODAI workshop (Workshop on Compilers, Deployment, and Tooling for AI) at HiPEAC 2026 (Kraków, Poland).

My talk, titled “The Compiler Before the Horse: Design Space Exploration at Fractile,” focused on how investing in compiler development early in our hardware design process has been essential for effective design space exploration and for making well-informed architectural decisions.

The stream was recorded, and is available from the CODAI website.

I tried to keep the tone light, and may have overdone it with my extended metaphor. Beating a dead horse, if you will.

In an age of easy, high-quality generative images, I had some fun old-school editing together my own images.

The first part of the talk mirrored a lot of the themes in my C4ML talk in Vegas in 2025, namely that we found a lot of success developing quick and dirty prototypes of the compiler using the MLIR Python bindings. We even got an end-to-end compilation and functional simulation of Llama2-7B, which was a great validation of our team, our design, and our high-level architecture.

I then discussed how the organisation has changed, and how as we approach tape-out, we have been nailing down the minutiae of our programming model, and how that interacts with our architecture.

The message I was trying to sell was that the compiler team, and staff at Fractile in general, have been empowered to make functional software early, to better understand the shape of our design space, and use that to inform our conversations with architects.

In this post, I chat with a former New Musical Express (NME) writer Glenn Gibson (aka, my da) about his experiences breaking into music journalism in the 1970s-80s, the evolution of music publications, the development of Scotland’s music scene during the punk and post-punk era, and various other bits and pieces.

You’ve got interesting stories about the music industry—I thought it’d be good to capture some of them. How did you get into music writing?

It was the New Musical Express, which was well-established when I started. It’s a music paper dating back to the 50s, and they were credited with originating the UK pop charts. But it was very much tied into the music business. They just had pop stuff and so forth. And at some point, during the early 1970s, they got a new editor who totally revamped it.

Starting in the 60s there was the underground press which was usually independent. And the NME started employing people who wrote for those publications.

So rather than just being a mouthpiece for what was coming out from the big labels, it was able to give space for some more up-and-coming and challenging artists?

Yes, rather than just covering what was in the charts. With the counterculture and all of the post-war developments, music was developing fast.

It went from rock and roll, to the Beatles, and the beat boom, and moved on to prog rock, glam rock, and punk. Around 1976/77 they wanted music publications that would reflect that — written by people who were close to the movement, who were really part of that.

They were probably copying what was happening in America because in the 60s, Rolling Stone magazine started. Rolling Stone is now well-established, I believe it’s quite a corporate thing now. But it started as a counterculture underground thing.

They employed the likes of Hunter S. Thompson for articles?

Yes, some real mavericks. One person they employed was Cameron Crowe, who went on to become a film director. A film he wrote, ‘Almost Famous,’ was semi-autobiographical.

Auch aye, I think we watched that together a few years back.

He was just somebody who was a big music fan. He sent them articles he’d written, and they liked them. I don’t think they found out until later that he was 15. They were sending him off to interview big-time stars and he became really successful. “Almost Famous” is arguably one of the best music business films ever made. I would say that and “24 Hour Party People” (2002).

And Cameron Crowe has actually just published his autobiography, “The Uncool” (2025), because part of “Almost Famous” was that he was never cool.

"That's when I started getting published—and paid."

Anyway, that’s the kind of thing that the NME were trying to copy, and it worked. They got a big audience and it became very cool, internationally as well. It was very well liked in America too. They seemed open to taking on people from around the country, so I went to see Ivor Cutler, the Scottish humorist, singer, songwriter, and poet, and I wrote something about it and sent it to them. They published it.

Oh, crackin’

They came back and said “would you like to do more?” I sent them quite a few things, and they didn’t publish any of them. But then the person in charge of the live review section changed and a new guy came in and he said, “Don’t do that anymore. What you should do is phone me every week and say what’s coming up and we will then commission you to write things.” That’s when I started getting published — and paid.

That being said, I still had a day job as well. Looking back, I don’t know how I managed it.

I guess not having weans helps.

I think the extreme would be working all day and then going to Edinburgh and not getting back home from Edinburgh until five in the morning. And then getting up to go to work. Out constantly at gigs and interviews, and also needing the time to write up.

Aye, I can imagine that took a good chunk of time. Did the acts you were sent to cover align with your tastes, or were you seeing a load of pish? Once they started commissioning, were they picking what you covered, or did you still have a say?

Well, a mixture, because I wasn’t doing exclusively what was happening in Scotland. I was being sent to review other bands who were touring who could be American, but as well as covering what was happening in Scotland because what was happening in Scotland was developing, and gaining national attention.

"One thing punk did was give people the message: 'you can do it too.'"

Starting to develop its own scene?

Yes. Simple Minds were one of the first ones. I went to see them playing in bars. I think they had already signed a contract with a record company at this point, so they were starting to move beyond that, but there were a lot of other ones coming up behind them.

There are probably quite a lot of bands that folk would be surprised to hear came out of the Scottish scene because the accent or dialect doesn’t necessarily come across.

Well, the music business was always very centred in London. This was after punk. One thing punk did was give people the message: “you can do it too.” And what they were playing was relatively simple. That simplicity had origins in early rock and roll. The British beat groups of the 1960s were actually quite ‘punk’ in their way—like The Who, The Animals, and Them. With “Them”, the singer was Van Morrison who is still very successful.

It was fairly simple, thus it’s something that people who were just teaching themselves how to play guitar or bass or drums could play. It was a reaction to progressive rock, which can be very complicated and long and was all about musicianship. That was a backlash against that. The impact of it was quite startling.

I guess there was starting to be more access to recording and distribution equipment during that time?

There were local recording studios, but access to recording didn’t change much during that period. With computers and software it became much easier to record from home, but you couldn’t do that then really, unless you bought quite expensive tape recorders, which could still be quite primitive. So bands would go to record in small Scottish studios. But it could still cost a lot of money. Some of them would be around £600 a day.

Auch, geezo, and I guess that’s £600 then as well.

For people without much money, that was a lot. And the Scottish bands who were playing in bars would be paid very little. I think in some cases they’d get like £10 and there might be four or five people in the band, and they had to hire a van to get there which might cost them about £70 — so they were making a loss. But they were just trying to get their own music out there.

It was an interesting dynamic but that message about “you can do it too” was quite widespread. And more than just music. For instance, my friend Martin Millar was very influenced by that ethos to become a writer. Actors, writers, people across the creative arts—a lot of people were influenced by that, and still are today.

I got to do [the writing] for a number of years and earned a reasonable amount of money and got to travel around and meet a lot of people. And with various different publications, I probably did it for about 11 years. Although you got some money, it was difficult to actually make a living.

Because it was quite sporadic?

Yeah, it was unreliable, I was just a freelance person who was also working a day job. There were some people who did the writing exclusively. However, they were probably getting unemployment benefits and things like that, or there might have been some other small local things that they did. There were a couple of people who went on to write for the local newspapers, doing the pop column.

Which is a bit more of a reliable income.

Yes, John Dingwall, for instance, he started off with a fanzine and then wrote for Record Mirror.

When he was just doing Record Mirror he probably wasn’t getting that much, but he went on to work for the Evening Times and Daily Record and he did that for a long time, and has been a core part of Scottish journalism since.

Making money from journalism at that time, even if you were employed by a Scottish newspaper which was predominantly sold in Scotland, you were paid quite well. The circulation of newspapers today has plummeted since then.

Aye, the issues that face the entire journalism industry today is a whole fraught topic unto itself.

I do know some people who moved to London and maybe had a bit of success. But they weren’t getting paid that much just working for music papers, or smaller music magazines, they were very often living in cheap accommodation: squats, somebody’s spare room, or flats with only cold water and broken windows.

I guess in a sense that’s quite punk. You’re getting an appreciation of your source material.

Right, the other side of that, is that the several people I can think of that did that, they all died quite young. In their early 60s and some even younger.

"They'd get like £10 and there might be four or five people in the band, and they had to hire a van which might cost them about £70—so they were making a loss."

Is that what a lifetime of substandard heating will do to ye?

Yeah, and maybe not being able to feed themselves properly as well. It might be a coincidence.

People say the body keeps score.

Yes, I obviously have nothing definite about it but you know, it happened to one photographer I worked with. He moved to London and got a job working for “Sounds” which was quite a good magazine. Quite well regarded. And I think after that, he even got a job working for Marvel.

I used to go and stay with him sometimes in London, but I kind of lost track after a while. Anyway, he got a brain aneurysm. He came back to Scotland a number of years ago and I think he’s in a care home.

The singer from the Bluebells, Ken McCluskey, who we knew, discovered that this guy Harry Papadopoulos was back in Scotland. And the reason Ken found out about it, was because Harry’s brother was an electrician or a tradesman of some kind and went and did a job in his house. The subject must have come up. Ken went to visit him, and they managed to get a hold of his old negatives. They’ve been on display in several art galleries and so forth. And there was a book published.

There was the “Rip It Up: The Story of Scottish Pop” exhibition at the National Museum of Scotland in 2018, where there were a couple of things you contributed.

I never really collected a lot of memorabilia, but there were some things I had. Part of the reason for it is that a lot of people in the past have been asked to lend things of that type and never got them back. So a lot of people who were asked to lend things would say “nah, not doing that”. But with the National Museum of Scotland they were known for being quite reliable, and you did get everything back. So it ended up being quite extensive, and showing off a lot of items never seen before.

Speaking of Scottish pop history being museum-worthy, what place do you think this sort of thing should have in our collective consciousness?

Occasionally in YouGov questionnaires they’ll ask things like “What does Britain do best?” I’m not saying things like industrial products—I’d say music, comedy, acting, writing, and the arts in general. And it’s hard to think of things that Britain does better than those kinds of things.

Aye, a lump of steel is more-or-less the same wherever it’s from. But music, comedy—that’s a product of the people and the culture. You can’t get that anywhere else.

For a very long time pop and rock music that was listened to internationally was almost exclusively from Britain and America. And that has been changing. There were certain countries, like Germany, that became quite successful, or at least critically recognised for what they were doing.

Aye, they weren’t playing Kraftwerk on Clyde One initially.

No. But there were a lot of other things coming out of Germany at that time, but it probably remained quite underground, or a niche interest. That’s continued—music from other countries gaining recognition in the UK. And now that you have K-pop and J-pop, these are massively successful.

Aye, it’s good to see that things are getting a bit more nuanced than just the catch-all term of “World Music”.

Yes, there were phases—African music was a thing for a while, and reggae especially — Jamaica was close to America which may have helped with discovery and distribution. It very much had its own sound, developing over the years from ska to blue beat to reggae, and they had their own version of rap quite early on.

Aye, and you see how dub influenced music worldwide. I’ve been noticing Afro-house having a similar reach lately.

Africa has so many different countries with their own traditions and own styles, but certainly around that time they would have had a lot of influence from what was coming out of Britain and America.

I suppose folk in Britain and America had a lot more access to the mass production and distribution required to get their music heard internationally.

A lot of these things are very popular in their own country and only certain ones break through internationally. There are certainly people that can help give it a boost. Paul Simon recorded an album with African musicians in the “Graceland” album a number of years ago. And he got a lot of criticism because South Africa had apartheid going on and there were supposed to be cultural boycotts.

However, I think his point was more that “this is the cultural bit and we’re actually paying people quite a lot more than the going rate”; and some of them got successful careers out of it internationally.

I wonder how many of those musicians were hitting oot with ‘I love apartheid, now here’s Wonderwall.’ Probably not fans of the regime themselves.

Most of the South African musicians involved were Black artists who had lived under apartheid and whose careers were constrained by it — but I think the boycott of South Africa was for the whole country. Therefore, even if you were working with people who were victims of it, it was still frowned upon in many ways. That being said, I think Paul Simon was quite a tough guy with a lot of self-belief, so he just did it anyway.

This was all well before my time, did it seem like the boycott was a significant part of how the apartheid came to an end?

Probably yes, but there are lots of factors beyond that.

Well, thanks for this chat—there are plenty of other threads we could pull on. The Virgin Megastores opening, people you’ve met over the years. Whenever a random track comes up you seem to have an anecdote ready.

I think for certain things I have quite a good memory. There’s this common thing that for some people, there’s certain types of music that they liked when they were in their late teens and that’s what they like forever. Whereas I’ve always wanted to find new things; and there are some people like that, but I think that’s the minority because you have to put some effort into it, and you do have to be interested. And if you’re interested enough you will remember these things.

I think to this day you’re my main source of interesting new music. I find stuff on my own, but it’s more through happenstance than actively seeking it out.

I think if you’re interested in music quite a lot, it ties in with so many other things. That would lead you into particular types of books or films or art. It’s all tied together.

Cheers da!

I was honoured to be invited to give a talk at the Alces Flight 2025 conference, which took place in London, UK, in October 2025.

My talk discussed how interdisciplinary collaboration and playfulness can be powerful tools, especially for early-career researchers and engineers.

I shared some of my experiences from my PhD and industry work.

At work, we recently transitioned everyone to MacBooks. I’ve never used one before, and had spent the past decade on Linux laptops which I’ve gradually refined to my liking.

Here’s an inconvenient truth — these machines are not designed to meet my needs. Being able to modify them to meet those needs has been a challenge. In an attempt to not alienate my readers, this doesn’t mean that folk with other needs are wrong. Apple clearly has a particular set of users in mind; I’m just not one of them.

My philosophy is that the tools I use should behave the way I tell them to. This post describes some of the issues I’ve encountered trying to achieve that, and how I’ve worked around them. Overall, I’ve been able to get a workable setup.

Windows and Workspace Management

One of the most important aspects of my workflow is the ability to manage windows and workspaces efficiently.

I typically have 10 workspaces, with a relatively strict purpose for each one, and map each to a keyboard shortcut, usually super+[0-9]. That way I can quickly switch between different tasks and contexts (e.g., text editor is in super+1, email in super+9, etc). I can also move windows between workspaces using super+shift+[0-9].

macOS has a concept of ‘Spaces’, which are virtual desktops, but the behaviour I required of sending windows to specific spaces and switching between them with keyboard shortcuts wasn’t available out of the box. I had to install a third-party window manager called yabai.

Yabai is a tiling window manager, which means that it automatically arranges windows in a grid-like layout. This isn’t something I’ve bothered with before, but I’ve gotten used to it, and I’ve ended up liking it. I’ve even enabled tiling windows on my Debian laptop, through KDE Plasma. It also helps to deal with the limited memory of the MacBook, since it pushes me to not have too many windows open at once.

Alas, Yabai and some other tools had some issues when the macOS Tahoe update came out in September 2025. This initially broke my setup, and made me feel like there was a small-calibre bullet in my skull, as it made everything much harder to use. Some patches have been released since then, and things are mostly working now.

Example Yabai configuration:

#!/bin/bash

# Re-inject scripting addition if Dock restarts
yabai -m signal --add event=dock_did_restart action="sudo yabai --load-sa"
sudo yabai --load-sa

# Update workspace number in SketchyBar
yabai -m signal --add event=space_changed action="$HOME/.config/sketchybar/plugins/space_display.sh"

# Set default layout to tiling (bsp)
yabai -m config layout bsp

# Window padding and gap between tiles
yabai -m config top_padding         0
yabai -m config bottom_padding      0
yabai -m config left_padding        0
yabai -m config right_padding       0
yabai -m config window_gap          0

# Focus behavior
yabai -m config mouse_follows_focus on
yabai -m config focus_follows_mouse autofocus

# Automatically rebalance split tree when windows are closed
yabai -m config auto_balance on

# Warp new windows into the tree (default tiling behavior)
yabai -m config window_placement second_child

# Enable window shadows and blur for floating windows (optional)
yabai -m config window_shadow on

# Make full-screen toggle (green button) act like regular zoom
yabai -m config window_opacity off

# Optional: float certain system apps (you can add more)
yabai -m rule --add app="System Preferences" manage=off
yabai -m rule --add app="System Settings" manage=off
yabai -m rule --add app="Software Update" manage=off
yabai -m rule --add app="App Store" manage=off
yabai -m rule --add title="Preferences" manage=off

# Optional: make the Dock not steal space if it's visible
yabai -m config window_shadow               off
yabai -m config window_opacity              off

You may wish to add spacing between windows, but I prefer to have none.

Keyboard shortcuts

I used skhd to manage the keyboard shortcuts for Yabai.

For example,

ctrl - 1 : yabai -m space --focus 1

ctrl - m : yabai -m window --toggle zoom-parent
ctrl + shift - c : yabai -m window --close

ctrl + shift - 1 : yabai -m window --space 1

I became frustrated with additional Mac keyboard issues and made further changes.

The Keyboard

The Mac keyboard is a bit different from a standard PC keyboard. And the keycaps are labelled differently. That layout doesn’t fit my needs, so I’ve had to make some changes.

First thing I did was remap the modifier keys in System Settings -> Keyboard -> Modifier Keys. I set Caps Lock to Command (to avoid Emacs pinky), and Command to Option (aka Alt). And the Option to Control (aka the Super key). You may want to swap the Command and Control keys instead, if you’re used to the classic Ctrl-c and Ctrl-v shortcuts for copy and paste.

The Mac keyboard is also missing a delete key, so I used an app called Karabiner-Elements to map Shift+Backspace to Delete. It has a GUI for remapping, and also has a JSON config file.

So for that example we can have:

{
    "profiles": [
        {
            "complex_modifications": {
                "rules": [
                    {
                        "description": "Shift + Delete = Forward Delete",
                        "manipulators": [
                            {
                                "from": {
                                    "key_code": "delete_or_backspace",
                                    "modifiers": { "mandatory": ["shift"] }
                                },
                                "to": [{ "key_code": "delete_forward" }],
                                "type": "basic"
                            }
                        ]
                    },

Other remaps worth considering include tilde (~), quote ("), at (@), and backslash (\).

External Monitor Issues

One of the most surprising issues was connecting to an external monitor. I like to connect the laptop to an external monitor, and I want the external monitor to be my only display (i.e., turn off the laptop screen). One monitor is sufficient for my workflow. It turns out, for some reason, this isn’t possible in Mac out of the box.

One can use “clamshell mode”, where if you close the laptop lid, the external monitor will become the only display. However, this means that you cannot use the laptop keyboard, and I prefer to use the laptop keyboard for consistency.

I had to install third-party software to achieve this, called BetterDisplay. In its settings, I then navigated to Displays->Overview->Connection Management Settings->Disconnect built-in display when an external display is connected. This capability should be built into the OS; it exists in most other systems.

SketchyBar

As part of using Yabai, I wanted a status bar that could show me the current workspace I’m using. However, out of the box, I couldn’t find a way to do this.

Therefore I used a third-party status bar called SketchyBar, which is highly customizable.

I opted to keep it simple, with the current workspace number, to the right of the battery level.

The default macOS status bar remains accessible with this configuration:

sketchybar --add item hide left \
           --set hide script="sketchybar --bar hidden=on; sleep 5; sketchybar --bar hidden=off" \
           --subscribe hide mouse.entered.global

But I’m grateful that SketchyBar exists, since there’s a lot of unnecessary menus and icons in the default Mac status bar.

Hiding the Mouse Pointer

I haven’t fully reached the point of never using the mouse, especially when much of the web is still designed for mouse interaction. However, I tend to use the mouse a lot less than most people. Therefore, I want the mouse pointer to disappear when I’m not using it. To achieve this, I used a third-party app called cursorcerer.

Emacs

The Emacs Plus formula from Homebrew is pretty good, but it doesn’t install a .app file into /Applications by default.

I had to copy the .app file into /Applications for the shortcuts to work.

sudo cp -R /opt/homebrew/opt/emacs-plus/Emacs.app /Applications

I also had to add (add-to-list 'default-frame-alist '(undecorated . t)) to my ~/.config/emacs/early-init.el file to remove the title bar, and rounded corners, which I consider unnecessary.

Terminal

Alas, I’m not so deep in the Emacs bucket that I use it as my terminal emulator. I prefer to use a separate terminal emulator, with tabs, and an extra level of nesting with tmux.

The default terminal on Mac had strange keybindings, and the app iTerm2 had similar issues.

I ended up using kitty, which allowed me to configure the keybindings to match my needs with a simple config file.

Conclusion

There are many more issues I’ve encountered, but these are the main ones that have got in the way of my workflow. There is a large community of third-party tools that may be better suited to meet the needs of your workflow, and I’m sure folk out there have even better setups. For me, this was the minimum viable setup to get my work done.

On the bright side, the experience of setting up a new environment has allowed me to feed back some improvements into my normal workflow on Linux as well. For example, Shift+Backspace to Delete, and tiling window management. Also, missing the Trackpoint on the Mac keyboard has pushed me to use the keyboard more, which feels like a win (in the most meaningless sense).

I’ve been chatting recently with Melize Ferrus, currently working on PhD in physics and studying various aspects of gravitational waves.

It’s not an area I have much expertise in, but through a series of chats and playing around with some of code (and making some fun demos), I thought I’d discuss what I’ve learned. This is not a deep dive into the physics, but rather an exercise in cross-discipline communication, and a chance to play with some Python code. I wrap up with a Q&A to hear take on physics and collaboration.

I’ve asked Melize to write some clarifications and commentary, which will appear in block quotes like this:

What it do, what it does 🥘

The rest of the text is my understanding, and may contain errors. Anything you see that is wrong is my fault, not Melize’s (including quote transcription). Everything should be prefixed with “I reckon”. You can read more about Melize’s work on gravitational waves in this article.

Melize’s code is closed source for now while works on some active areas of research, but I was given access to it for this exercise. The astropy and numpy libraries are used extensively.

The Problem Area

The code that Melize shared with me was looking at binary white dwarf (BWD) systems. These are pairs of white dwarf stars (which are the cores of dead stars) orbiting each other.

There are a few reasons why these systems are interesting to study. From what I gather though, of particular interest is that when these stars get close to each other, they can produce very small gravitational waves.

When the white dwarfs get close enough that they become bound and start orbiting each other, they emit gravitational waves. We’re mostly looking at ones inside our own galaxy, because anything further out would be way too faint to detect.

It is difficult to detect gravitational waves from such small systems, but the reason why we care is that they contribute noise, or “background radiation”, to the gravitational wave signals we do care about, such as those from black hole mergers.

Therefore, understanding these white dwarf systems helps us to better interpret the gravitational wave data we receive from detectors like LIGO and Virgo.

In my case the work’s for LISA — the space-based gravitational wave detector. It’ll be much more sensitive and looks at a totally different frequency range than the ground ones like LIGO or Virgo. You need it in space, because the arms have to be effectively millions of kilometers long. Ground detectors catch signals lasting less than a second; LISA will see them for weeks or even months. That means we can detect black holes long before they merge and coordinate with telescopes to watch the same event. It’s going to open up a lot for cosmology and multi-messenger studies.

So how does Melize’s code tackle this problem?

The Code

I saw that Melize’s code had two main parts, both of which use Gibbs Sampling. Gibbs sampling is a statistical technique used to generate samples from a probability distribution. We don’t need to use “real” data points from astronomical observations, especially with the number of data points we need. Instead, we can use our models of the Galaxy and the binary white dwarf systems to generate synthetic data points that are statistically similar to what we might observe.

I’d maybe mention it’s a Monte Carlo method — Gibbs sampling is just one type of that. There are a bunch of approaches; this is the one I happened to use.

The two main uses of Gibbs sampling in code were:

• Drawing samples of what the binary white dwarf systems might look like (e.g., their masses, distances, and orbital periods).
• Generating where in the galaxy these systems might be located.

Boffins have modelled what we think the distribution of these systems looks like, and this code is drawing samples from this distribution. The specifics of the distributions are not important for this discussion, and not something I’ve taken the time to understand.

It’s only separated into two parts because I’m lazy. Theoretically they could all live in the same Gibbs sampler. It’s on my to-do list to merge them into a proper joint distribution.

Binary White Dwarf Systems

When we generate a sample of a binary white dwarf system, we are generating:

• The mass of the first white dwarf (in solar masses).
• The mass of the second white dwarf (in solar masses).
• Their semi-major axis (in solar radii) — essentially the orbital radius, but these orbits aren’t circular.
• Their orbital eccentricity (what shape their orbit is).

Semi-major axis is the halfway point of the furthest distance between the two bodies in their orbit.

For example, a sample I generated from Melize’s code was:

Example Binary White Dwarf System

\(\begin{aligned} m_1 &= 1.33\,M_\odot \\ m_2 &= 0.64\,M_\odot \\ e &= 0.38 \\ a &= 7.26 \times 10^5\,R_\odot \end{aligned}\)

To help with my own exploration, I also gave each system a random name using 2 words from the EFF Long Wordlist.

E.g.,

overrun-nicotine
stack-thursday
goldmine-undesired
strained-hurried
ecologist-undead
cherisher-degree
supermom-unsigned
rescuer-sternum
zebra-crepe

From this code, we can generate as many samples as we want, and our set of binary white dwarf systems will be distributed according to our model.

Melize, what can we do with these samples?

What I’m actually computing is the gravitational-wave power contributed by all the white dwarfs at each discrete frequency. E.g., using some nonsense numbers and units, it might give one million watts at 1 mHz and something else at 10 mHz. Let’s say those are the bands LISA’s interferometer is looking at, and that helps us correct our observations. That’s my focus, though you could use the samples for other analyses too.

Galactic Coordinates

We expect the white dwarf systems to be distributed around the Milky Way galaxy, with a higher density towards the center of the galaxy, and a lower density as we move away from the center.

The galaxy is modelled as a disk, with a bulge in the center. Melize’s model generates “2D cylindrical galactic coordinates”, which are:

• (R): radial distance: the distance from the center of the galaxy.
• (Z): the vertical distance above or below the galactic midplane.

The coordinates are cylindrical — the center of the galaxy’s at (0, 0). They’re distributed to match how main-sequence stars are spread throughout the Milky Way. From a logical standpoint they’re just dead stars, so BWDs will be wherever stars are. I can’t think of a strong reason for their distribution to be very different.

This is in contrast to the more familiar Cartesian coordinates (x, y, z). However, to convert from galactic coordinates to Cartesian coordinates, we also need the “azimuthal” angle (φ), which is the angle around the center of the galaxy. This requires an assumption about how the systems are distributed around the galaxy. This isn’t necessary for Melize’s work, but I wanted to visualise the results in a format I was more familiar with.

Therefore I added a third step to the code, which was to draw a random angle between [0, τ], and then convert from galactic coordinates to Cartesian ones. We already have the Z coordinate, and we can convert from (R, φ) to (x, y) using:

\(\begin{aligned} x &= R \cos(\phi) \\ y &= R \sin(\phi) \end{aligned}\)

For what I’m doing, you do eventually need the azimuthal angle because you need to know where the binary is located relative to the gravitational wave. But it’s uniform, so there’s no need to sample it — it’s just drawn uniformly, sometimes conditional on other factors.

Visualisation

With Melize’s code, and some minor data transformation, we can now generate 5,000 BWD systems, and also generate their positions in the galaxy. The code is capable of generating as many samples as we want, but for my purposes I kept the number small.

This plot is interactive, using the plotly library, and its go.Scatter3d tool. I initially made this in a Jupyter notebook, and then exported it to a format that can run in the browser using Plotly’s JavaScript library, using fig.write_json("2025-10-binaries.json").

# Create an interactive 3D scatter plot of the sampled systems
import plotly.graph_objects as go

fig = go.Figure(
    go.Scatter3d(
        x=x,
        y=y,
        z=z_vals,
        mode="markers",
        text=hover_text,
        hoverinfo="text",
        marker=dict(
            size=3,
            color=m1_vals,
            colorscale="Turbo",
            colorbar=dict(title="Primary Mass [M☉]", tickcolor="white"),
            opacity=0.8,
        ),
    )
)

You can see it below, if your browser supports it.

If you hover over a point, it will show you the details of that binary white dwarf system, including the randomly generated name I added.

Audio Shenanigans

Next, I wanted to do something more playful. Using the generated BWD systems, I tried out a few techniques to turn them into musical notes. I spent some time tuning how different parameters influenced the audio.

I’ve not done much audio programming before, so I kept it simple, using 4 basic waveform types: sine, square, snare, and kick.

I generated a bunch of BWD samples, and checked the parameter ranges so I could normalise them.

My code ended up being quite hacky:

def bwd_to_voice(m1, m2, e, a, norm_ranges):
    """Map BWD parameters to sound features and instrument type."""
    nm1, nm2, ne, na = norm_ranges
    mass_avg = (m1 - nm1[0]) / (nm1[1] - nm1[0] + 1e-9)
    ecc = (e - ne[0]) / (ne[1] - ne[0] + 1e-9)
    sep = (a - na[0]) / (na[1] - na[0] + 1e-9)
    q = m2 / m1
    asym = abs(m1 - m2) / m1

    # instrument choice
    if mass_avg > 0.8:
        voice = "kick"
    elif ecc > 0.7:
        voice = "snare"
    elif asym > 0.3:
        voice = "square"
    else:
        voice = "sine"

    freq = 100 + 800 * mass_avg
    dur = 0.2 + 0.8 * (1 - sep)
    vol = 0.3 + 0.7 * ecc
    return freq, dur, vol, voice

With a constant BPM, I passed binary white dwarf systems through this code, making one note each, and generated the following track:

A tad atonal, which makes sense since we’re mapping astronomical parameters to musical ones without any harmonic constraints. Maybe I need to read more music theory. However, it grows on me the more I generate and listen to it.

Perhaps it’s hubris to think that randomly generated white dwarf systems would immediately make good music, but it’s fun to try.

I can’t wait to somehow contribute to a future post where we come at music from the mathematical angle — figuring out what makes music sound good without actually knowing what the notes mean.

Q&A

Now that I’ve shown some of my hacking around with Melize’s code, here I take the opportunity to ask some general questions.

What got you interested in this area of research?

Honestly, nothing in particular. I always wanted to be an astronomer as a kid. At some point I realized I liked physics, so I thought astrophysics would be cool.

But I didn’t really know how to read my college website properly, so I didn’t realize they didn’t even have astronomy. I knew they had physics, but I didn’t check the course list well enough. Once I was there, I realized—oh, there’s nothing astronomy-related here.

I got involved in aerospace engineering stuff for a bit, then kind of gave up and moved on. Later, I heard about an internship at the Flatiron Institute for people in the National Society for Black Physicists. I applied, got sent a list of different research topics, and picked my top three. The one I ended up doing was literally my lowest-ranked choice.

I think my top pick was something about solar oscillations, and the second one I forget. But yeah, I ended up getting the third one, and over the course of the internship I realized I really liked it. So there was no grand plan like, “I’ve always wanted to work on black holes.” It just kind of happened.

What are some big open questions in this area of research that you’re interested in?

I work across a pretty wide range of things in gravitational waves. What I’m doing now is different from what I did with at Flatiron, and different again from what I did at King’s.

On the numerical relativity side, one big question is whether we can come up with a model or theorem that works for the merger phase of black hole collisions. We already have post-Newtonian approximations for the inspiral, and quasi-normal modes for the ringdown. Those let you predict what the gravitational wave looks like if you know the parameters of the system.

But for that last bit, right before the black holes coalesce, there’s no theorem—just heavy computation. It would be great if we had a simpler equation or model instead of running six-hour-long simulations for that one stage.

Then there’s eccentricity. Black hole mergers can have different eccentricities, and that changes the whole signal. But those effects can look like other things—spin, non-vacuum conditions, whatever. So it’s hard to tell what’s what. We usually assume circular orbits, but they’re often not. Understanding the typical black hole orbit better would help a lot—both for improving models and for things like star formation studies.

For the LISA mission, a huge open question is what we’ll learn about cosmology. LIGO only catches the last second or so of a merger. LISA will detect black holes a month in advance. That means you can get detailed parameters early, subtract out the known signals, and study what’s left—possibly signs of inflation, cosmic strings, or first-order phase transitions. The future looks really exciting for cosmology.

You deal with some pretty large simulations and computations. Would you say that programming is a big part of your work? Is there enough collaboration between physicists and computer experts?

Yeah, programming’s a big part of it. But I’d say most physicists are computer experts only in the sense that they know enough to make their own stuff work. We’re not computer experts broadly.

Like, you [ed: Perry] know enough physics to help me out with something, but not enough physics overall to anticipate what I might need. Same for me—I know enough computational stuff to get results, but I don’t always follow the best practices. Someone else could look at my code and think, “Sure, it works, but this could be done way better.”

There could definitely be more collaboration. Everyone’s under pressure to produce results quickly, so code often ends up written in brute-force, opaque ways. Later, when new people join, they spend weeks just figuring out what the code even does. Half of it could’ve been explained in a docstring or wiki.

A lot of scientific code isn’t open source. That really limits reproducibility and collaboration. Sometimes groups hoard code to keep control—people have to email them to run a simulation instead of doing it themselves. Then those teams rack up citations for just pressing run. Nothing’s truly replicable because others can’t see the underlying code.

I’ve even been on papers where the authors said they were using the “highly precise [system] code.” But how do you know it’s precise? You’ve never seen the infrastructure. Everyone assumes it’s right because a big name’s attached. It’s very anti-scientific—this blind trust that something’s correct just because of who made it.

Amongst other things, you’re black, and present pretty femme, which is still relatively rare in physics. Have you had any interesting or noteworthy experiences through this lens?

I’ve got a few specific examples, but I don’t want to call anyone out.

I get mistaken for other people a lot, even at conferences. It happens so often it’s almost funny, except it’s not.

Like, once at a conference, I asked in the group chat if anyone had painkillers because I had cramps. A guy brought me one, fine. The next day, he said he’d seen “me” while biking and shouted, “Do you still need the drugs?”—but it was just some random woman with similar hair. He figured that must be me.

I’ve also been to plenty of conferences where no one sits with me. Literally, people will fill every seat around me but not next to me. It feels isolating.

It’s not that anyone’s openly rude. It’s just that constant, quiet exclusion. Seeing your peers welcomed while you’re sidelined makes you ask why. It’s tiring.

What are the biggest uncertainties in the models you work with?

Oh, huge ones. One of the biggest is the merger rate—how many mergers happen per year. The uncertainty can be as large as the value itself.

Let me bring up one of my favorite ones. It’s very entertaining to me

( types on laptop for a moment, then starts laughing.)

Here we go: neutron star mergers [page 26]. The estimated rate is 920 per gigaparsec cubed per year, with a +2,220 / –790 margin!

Hands down, that’s my favourite. I roll around in my non-existent grave to this shit. It’s a couple of years old, so newer estimates are probably tighter now. This is for the first two observations of LIGO with the neutron star detection for the collision rate.

Yeah, I have a lot of confidence in all of the different things I generate using those kinds of things…

Anything else you want to say?

I really like the kind of work I do. I’ve been lucky to move between institutions and completely change up what I’m working on each time.

After my year at Flatiron I felt cocky about starting my PhD, “I run this bitch,” but when I got to King’s it was like starting from scratch again due to the alternative perspectives on the topic. It’s cool seeing how much undergrad material—stuff I once thought was useless—actually shows up all the time.

In undergrad, you’re rushing to master something in 13 weeks, then it leaks out of your head. Doing a PhD gives you the time to actually learn it properly. None of my physics education was a waste. Even the quantum mechanics I thought I’d never need has ended up being useful.

It gives you time to learn the thing properly, as opposed to learn the thing in order to pass the test. Yeah, so like if you didn’t do everything perfectly in your undergrad, it’s fine. I think it’s just that doing a PhD gives you more appreciation of it.

Thank you Melize.

Thank you Dr Gibson, I’ll see you on GitHub 🙂‍↕️.

Conclusions

When I talk to specialists in other fields who are articulate and passionate about their work, I find it infectious. I usually come away with the feeling of “ah, so it isn’t magic, but it is hard”. I haven’t eaten my intellectual vegetables by working through general-relativity textbook exercises. However, by inhabiting Melize’s code for a while, and picking brain, I feel like I’ve gained a little more appreciation of the field, and the cool people that work in it.

I’ve recently been chatting with a couple of mates who are doing scientific research using Python. Answering specific code questions is a lot of fun, as I try to figure out enough of their domain to see if/how I can help.

However, there are some general workflow improvements that I think both they and other scientific Python programmers could benefit from. In this post, I’ll share some lightweight workflow improvements that can help make your code better.

Research code is often different from production software: it doesn’t need to be as robust or strictly engineered. However, a few simple practices can make it far more reproducible and maintainable.

In this post, we’ll explore steps to improve reproducibility, collaboration, and confidence in your results — without overcomplicating your code.

This post assumes you’re using GitHub (or a similar platform like GitLab or Gitea) for version control. If you’re not yet using a VCS, I highly recommend starting there.

If you are working on more software-engineering-oriented projects, this post is probably not for you, as there are many more things you should be doing.

We’ll go through a few simple steps:

• Creating a Python package with a reproducible environment.
• Writing tests with pytest to catch bugs early.
• Using pre-commit hooks to maintain code quality automatically.
• Setting up GitHub Actions to run tests on every push.

The Project

Let’s say we’re using Astropy to do some cosmology calculations. This isn’t my area of expertise, I’ve just made some bullshit code that shows the development practices: apologies if it offends your sensibilities!

Initially, we explored these calculations in a Jupyter notebook. While notebooks are great for experimentation, they aren’t ideal for sharing, testing, or reusing code. I’ve seen people with functions they copy between every notebook, making slight changes each time, and then lose track of which version is “the good one”.

Our first goal is to take some commonly used functions from the notebook and turn them into a small, reusable Python package that we can import into any notebook or script.

Common Package

The outcome of this section can be seen on commit ed941a of this GitHub repository. I’ll walk through the steps below.

When you have functions that you want to reuse across multiple notebooks or scripts, it’s best to put them in a common package rather than copying and pasting. Copying code can quickly lead to inconsistencies and bugs as your code evolves.

For this project, we’ll create a small package called cosmole containing three scripts:

.
├── cosmole
│   ├── __init__.py
│   ├── angular_separation.py
│   ├── convert_equatorial_to_galactic.py
│   └── redshift_to_distance.py
└── notebooks
    └── 2025-10-07-exploration.ipynb

The __init__.py file defines what is accessible when someone imports the package:

# cosmole/__init__.py
from .redshift_to_distance import redshift_to_distance
from .angular_separation import angular_separation
from .convert_equatorial_to_galactic import convert_equatorial_to_galactic

Now (once we’ve finished the following sections), from any notebook or script in the project, you can import your functions:

from cosmole import (
    redshift_to_distance,
    angular_separation,
    convert_equatorial_to_galactic,
)

Dependency Management

In scientific projects, it’s easy for code to stop working later if software packages you depend on change. When I was writing up my PhD, I had to regenerate some plots from my first year and ran into issues because matplotlib had changed; I wasn’t sure which version I had originally used — it was a pain.

To avoid “it worked on my machine” problems, we can specify exact package versions using a requirements.txt file.

This is a simple text file listing all the Python packages your project depends on, along with their versions. For our project, you can see it here.

You can install these dependencies with:

pip install -r requirements.txt

Or with conda:

conda install --file requirements.txt

This ensures anyone cloning the repository can reproduce your environment and run your code without surprises.

Setting up the package

To make Python aware of your cosmole package, we need a setup.py file at the root of the project. This file tells Python how to install your package and its dependencies. In our example project, you can see it here.

We can now install our package with:

pip install -e .

• The -e flag means editable — any changes you make to the source code are reflected immediately when you import the package.
• This is especially handy when working interactively in notebooks.

Additionally, adding and executing these lines at the top of your notebooks will also mean you automatically reload your package whenever you edit any of your .py files:

# Enable autoreload so that changes in imported modules are reflected automatically
%load_ext autoreload
%autoreload 2

Note to Conda users

If you use Conda, you might need an additional file pyproject.toml in your project to ensure that the package installs correctly.

E.g.,

[build-system]
requires = ["setuptools", "wheel"]
build-backend = "setuptools.build_meta"

This is because modern versions of pip build packages in isolated environments. So you need to specify that setuptools and wheel are required to build the package (which then installs the dependencies from requirements.txt).

Testing Your Code

The outcome of this section can be seen on commit 9905048.

Once you have refactored your functions into your package, it’s a good idea to write tests.

Actually, one approach is to write the tests first as we’re developing our initial implementation of our functions (a practice known as test-driven development). However, in research, this might not always be the most practical approach.

Tests help ensure that your code behaves as expected, which is especially useful when making changes or refactoring. Good test design is an art, thinking about edge cases and failure modes.

Even if your research code is exploratory, having a few basic automated tests can catch obvious bugs early and give you confidence that updates don’t break existing functionality.

During development, you might find that a test itself was wrong, or you changed the intended purpose of a function. In which case, just update the test — make them work for you!

We can use the pytest package to write and run tests. For example, let’s test one of our functions by adding a file tests/test_convert_equatorial_to_galactic.py to the project:

import pytest
from astropy.coordinates import SkyCoord
import astropy.units as u
from cosmole import convert_equatorial_to_galactic

def test_convert_equatorial_to_galactic():
    ra, dec = 83.82208, -5.39111  # Approx location of Orion Nebula
    l, b = convert_equatorial_to_galactic(ra, dec)
    coord = SkyCoord(ra=ra*u.deg, dec=dec*u.deg, frame='icrs').galactic
    assert pytest.approx(l, rel=1e-6) == coord.l.degree
    assert pytest.approx(b, rel=1e-6) == coord.b.degree

The assert pytest.approx lines verify that the function returns the expected results, raising an error if it doesn’t.

To run the tests, from your project root, run:

pytest

You should see output like this:

tests/test_angular_separation.py .                                                               [ 33%]
tests/test_convert_equatorial_to_galactic.py .                                                   [ 66%]
tests/test_redshift_to_distance.py .
[100%]

Which shows that all our tests passed. If we’d introduced a bug (e.g., I accidentally added 10 to one of my results), the test would catch this:

tests/test_convert_equatorial_to_galactic.py F                                                   [ 66%]
tests/test_redshift_to_distance.py .                                                             [100%]

=============================================== FAILURES ===============================================
_________________________________ test_convert_equatorial_to_galactic __________________________________

    def test_convert_equatorial_to_galactic():
        ra, dec = 83.82208, -5.39111  # Approx location of Orion Nebula
        l, b = convert_equatorial_to_galactic(ra, dec)
        coord = SkyCoord(ra=ra * u.deg, dec=dec * u.deg, frame="icrs").galactic
>       assert pytest.approx(l, rel=1e-6) == coord.l.degree
E       assert 219.01374318285238 ± 2.2e-04 == np.float64(209.01374318285238)
E
E         comparison failed
E         Obtained: 209.01374318285238
E         Expected: 219.01374318285238 ± 2.2e-04

tests/test_convert_equatorial_to_galactic.py:11: AssertionError

Pre-commit Hooks

To help maintain consistent code quality, we can use pre-commit hooks. These are scripts that automatically run before each commit to check for common issues like code formatting, linting, or missing documentation.

The key benefits of pre-commit hooks are:

• Catch issues early — errors or style violations are flagged before they enter your repository.
• Automatic fixes — many hooks, like code formatters, can fix problems automatically, saving time.
• Consistency across contributors — ensures everyone on the project follows the same style and quality rules.

Configuration

Pre-commit hooks are configured in a .pre-commit-config.yaml file at the root of your project. For example, a simple configuration might include:

• Ruff — a fast Python linter and code formatter.
• Prettier — for formatting YAML, Markdown, and JSON.

In our example project, you can see the configuration file here.

You can install pre-commit and set it up with:

pip install pre-commit
pre-commit install

Now, every time you make a commit, the hooks defined in your configuration file will run automatically on any files that have changed. You can also run them manually on all files with:

pre-commit run --all-files

In our code, several automatic reformatting steps were applied: none of which should change our functionality, only the neatness of our code. For example, removing unused imports, making sure lines don’t get too long, being consistent with the use of whitespace, etc.

However, there are also some issues that need manual review. For example, Ruff complained that we had an ambiguous variable name l (lowercase L - which looks like a 1).

tests/test_convert_equatorial_to_galactic.py:8:5: E741 Ambiguous variable name: `l`
   |
 6 | def test_convert_equatorial_to_galactic():
 7 |     ra, dec = 83.82208, -5.39111  # Approx location of Orion Nebula
 8 |     l, b = convert_equatorial_to_galactic(ra, dec)
   |     ^ E741
 9 |     coord = SkyCoord(ra=ra*u.deg, dec=dec*u.deg, frame='icrs').galactic
10 |     assert pytest.approx(l, rel=1e-6) == coord.l.degree
   |

Normally, I’d recommend following the rules — they’re opinionated but usually sensible.

However, we can tell Ruff to ignore this specific warning by adding a comment to the end of the line.

l, b = convert_equatorial_to_galactic(ra, dec)  # noqa: E741

You can skip pre-commit using --no-verify, but only do so if you have a good reason (like it’s 10pm and you just want to get your changes saved before bed 😴).

In my example, you can see all the pre-commit changes in d23d21a. In practice, these changes would be applied at the time of writing the code in the first place, but I wanted to show the before-and-after.

GitHub Actions

Now, we have tests and pre-commit hooks, which are great for local development. However, we also want to make sure that our code is tested when we push changes, especially if we’re collaborating with others.

GitHub Actions allows you to run workflows on GitHub-hosted machines. While these have limited compute power, they can easily handle running basic tests and automation tasks.

For our example, we can create a workflow that runs our tests and pre-commit hooks on every push to the repository. If your code can run and pass on a different machine, that is a great indicator that it is reproducible and not dependent on your local setup.

Let’s do this by creating a file .github/workflows/ci.yml as shown here.

Now, every time you push changes or create a pull request, GitHub Actions will automatically run your tests and hooks. This adds an extra layer of confidence and helps maintain a reliable, reproducible scientific codebase.

You can see on the repo’s GitHub page, a yellow circle after we’ve pushed, meaning the checks are running

You can see initially, this run fails, and we can click on the red ❌ to see the terminal logging what went wrong.

In this case, we didn’t actually install our package on the GitHub runner. Therefore, we need to add pip install . to the workflow:

- name: Install the package
  run: pip install .

Now, our code passes!

In future, we’ll be able to see which commits introduced bugs in our code. If you use a git feature branch workflow (using branches for features and PRs), then you can enforce that any commits to main must pass these checks. However, this may be overkill for a small research project.

Conclusion

Ultimately, the goal of research code is to help us explore and answer our research questions. However, small software engineering standards can pay dividends if applied in a sensible way.

There are more things you may want to explore, for example, you can add more rules to Ruff (like enforcing documentation). Similar workflows can also be applied to projects in other programming languages. You may also want to standardise your git commit message format, e.g., using Conventional Commits.

P.S. props to my black-hole buddy Melize for providing the header image, which is his reproduction of Korol et al. 2018.