Probabilistic modelling and statistical inference.
Machine learning models are probabilistic models. Treated as such they can be reasoned about: their uncertainties carry meaning, their failure modes become predictable, and the conclusions drawn from them can be defended. Treated as black boxes, none of that holds.
I work with research groups and companies on that boundary — putting machine learning inside statistical inference that is mathematically sound, and understood by the people who have to rely on the answer.
Neural networks, including the large language models behind most of what is currently called AI, are probabilistic models, almost always fitted via some form of maximum likelihood estimation. Treating the networks as distributions, rather than as machines that produce answers, makes them possible to interrogate: what distribution has a model actually learned, what are its outputs conditioned on, and what do its inferences mean?
However, in their current form, such networks are massively misunderstood and their results mistreated. A softmax score is not the probability that an answer is correct. An ensemble with low variance is not necessarily a well-calibrated one. A model that fits the data can still be the wrong model, and will fail confidently when it is. These are just a few misunderstandings among many, but the results are routinely and dangerously treated as concrete fact.
The difficulty compounds when a learned component is placed inside a model you already have — a physical simulation, a production or optimisation model, anything where you want to understand the answer and not simply obtain a value. Much of my work sits there: keeping the physical effects interpretable once a data-driven part sits among them, and establishing how far the uncertainty on the answer can be characterised at all. Knowing when it can and cannot be, and why, is the more valuable result.
Projects have covered climate and weather modelling, heating and cooling systems, energy storage and grid management, the CO2 footprint of steel manufacture, and process and production optimisation.
Client names are omitted under NDA.
Engagements run from a single workshop to long-term working relationships. Either way, your team should end up owning and understanding what we build: able to carry it on without me, and to bring me back because it is worth doing rather than because they are stuck.
Open tools by default. What I build with you should keep working once the engagement ends, without a licence, a subscription or an account.
Published work dates mostly from my time in academia, with occasional co-authorship since. Papers on arXiv.
tom@charnock.fr
Based in France, working remotely worldwide.