About Me

I am an independent AI and machine-learning consultant based in Paris, working remotely with clients across Europe. I build AI systems for production — RAG, agents and LLM pipelines — and forecasting models where a decision depends on the number. I hold a PhD in computer science from INSA Lyon, carried out at the LIRIS laboratory (CNRS, UMR 5205) under Professors Céline Robardet and Marc Plantevit, on machine learning for electricity price forecasting on the European day-ahead market. It was a CIFRE industrial thesis with BCM Energy (then trading as Planète Oui, now Elmy), defended on 17 January 2024.

The research, and why it is here

A doctorate is worth stating for one reason: it is where the habits came from. Four years of being wrong in public, where a bad forecast cost the desk money the same afternoon, teaches validation design better than tidy benchmark work does.

The thesis, Machine learning techniques for electricity price forecasting, predicts the twenty-four hourly prices of a European day-ahead market before they are fixed at noon. The literature splits into two incomplete camps. Expert models replicate the price-fixing algorithm from estimated inputs: theoretically right, inaccurate in practice, because input errors propagate through a clearing process never designed to be robust to them. Data-driven models estimate prices directly: accurate, and opaque. Opacity is not an abstract complaint when the output goes to a trading desk — a trader who cannot see why a forecast moved will not size a position on it. Explainability is an adoption requirement, not a compliance nicety: the most portable lesson in the thesis.

Read the thesis (PDF) · defence slides · defence video

Putting an optimisation problem inside a model

European day-ahead prices are not observed, they are computed: a single algorithm — EUPHEMIA — clears every submitted order at once, maximising social welfare subject to energy balance and cross-border capacity. Almost every forecasting model in the literature treats the market as a black box instead. My DSAA 2023 paper does the opposite: it predicts the order book from fundamental variables, then feeds those orders into a differentiable formulation of the clearing problem, so the mechanism is present during training rather than bolted on at the end. That required a differentiable dichotomy search solving a mixed-integer quadratic program inside a neural network.

That shape is not confined to electricity. Whenever the quantity you are forecasting is itself the output of a known process — a pricing rule, an allocation, an auction, a scheduler — the same three options apply: ignore it, replicate it and inherit its input sensitivity, or make it differentiable and train through it.

Publications

Three peer-reviewed papers and a thesis. 329 citations as of September 2026 (Google Scholar).

Frontier Prize — best paper, IDA 2023. Forecasting Electricity Prices: An Optimize Then Predict-Based Approach. Intelligent Data Analysis 2023, Louvain-la-Neuve. Springer LNCS 13876. doi:10.1007/978-3-031-30047-9_35
Electricity price forecasting on the day-ahead market using machine learning. Applied Energy 313, 118752, Elsevier, May 2022. The most cited. doi:10.1016/j.apenergy.2022.118752
Electricity Price Forecasting based on Order Books: a differentiable optimization approach. IEEE DSAA 2023, Thessaloniki, with Tias Guns. doi:10.1109/DSAA60987.2023.10302542
Machine learning techniques for electricity price forecasting. PhD thesis, INSA de Lyon, January 2024, number 2024ISAL0007.

Education

PhD, Computer Science — INSA Lyon, LIRIS (CNRS), 2020–2024.
Diplôme d'ingénieur, Computer Science — INSA Rouen, 2014–2019, Information Systems Architect specialisation, Master's level (EQF 7). Semester at The University of Queensland, Brisbane.

Curriculum vitae (PDF). French (native), English (professional).

Working together

A small number of clients at a time, two to three days a week, remote from Paris and on site in France when the work needs it. If you have an AI prototype that has to become a system, or a forecast that has to drive a decision, book thirty minutes.