Skip to main content

Reflection

How I saw the role, and how I see it now

I discovered machine learning in 2018 through online tutorials, and what I remember is that it felt like magic. That feeling came back when ChatGPT was released, and it left me with a conviction I now consider wrong: that large language models would absorb everything else, and that classical machine learning would fade out. Interviews with companies, and the training itself, corrected that. These families of tools serve different purposes, they are complementary, and they will coexist. The larger shift is elsewhere. I started out interested in models and ended up interested in the system around the model: structuring it, securing it, making it operable in production and economically viable. My path also narrowed by elimination rather than by attraction. P12 showed me that data engineering is not my axis, and P09, a project without a single line of code, turned out to be the closest to the role I want. What I took from my first engineering degree in the wood industry proved directly transferable: structure the thinking, measure before and after, then propose.

What I would do differently

Today every project runs under an explicit method: what I must master and write myself, what I need to understand but not produce from memory, and what is plumbing that I take as is. It took two failures to get there. I spent the first months trying to learn everything that was asked, in full, and the time budget exploded. I then swung to the opposite: the code was produced for me and I transcribed it. The backlog shrank, the learning did not.

The second thing I would change is older. When I built the ASR pipeline in 2025, I evaluated the correction dictionary on the data that had been used to build it. I did not see the circularity at the time. I would now define a held out test set before writing anything, which is exactly what my current project does. I would also have run a feasibility spike on that pipeline before committing to it, to check it was achievable for me and for my colleagues. It worked, so I got lucky. It could have consumed weeks before I found out.

Progress I can already point to

Two are technical. I can build a real retrieval system rather than a large context stuffed with documentation, and more importantly I can measure whether it answers correctly, with Ragas and with the sources displayed next to each answer. And I write Python: before the training, the only Python I produced came from vibe coding.

The one that matters more is a change of judgement. I used to stop at a metric that looked good. On P14, fine tuning a model for hospital triage, the training metrics converged while clinical accuracy degraded. A metric never certifies an output on its own. That is why the roadmap I proposed for that project put human verification on the outputs rather than more monitoring on the metrics.

What is still ahead

Three things. I have never delivered an AI project for a client under an imposed deadline and budget, even though I have consulting experience in other fields. I have trained models on cloud infrastructure but I have never operated a service over time, with monitoring, alerts and incidents. And I will work alongside far more experienced developers, as someone from the generation whose code is partly produced by coding assistants, facing people who wrote everything by hand for years. I do not yet know exactly what I produce without assistance. Finding that out is part of what comes next.