Projects

A Python framework, built from scratch, that models the firing dynamics of prefrontal cortex populations with a system of ODEs: a 4-population rate model of the PFC microcircuit, extended with a ring attractor network for spatial working memory.

The heart of the project is the fitting side. I wrote a custom optimization pipeline (built on Nevergrad) to fit the model to real biological recordings. The core difficulty is that a plausible fit is not enough: the parameters have to produce a genuinely bistable network that matches the data in both the high and the low firing-rate states at once. Most of the engineering went into shaping the loss and search so the optimizer converges on that bistable regime rather than on a single-state solution that scores well on the average.

The model is then used to study how disruption of nicotinic acetylcholine receptors (α7, β2 and α5 subtypes) alters PFC dynamics in the context of Alzheimer’s disease.

The codebase ships a full CLI, a test suite, and extensive mathematical documentation covering Wong-Wang transfer functions, NMDA gating variables, and spike-frequency adaptation. Built during my master’s thesis at the Group of Neural Theory, ENS Paris.

Stack: Python, Jupyter Notebook, Nevergrad, NumPy, Matplotlib

A follow-up to my thesis work. Fitting a bistable network was the hardest part of AD_pfc_modeling, and I solved it with a hand-designed loss. This project asks whether an LLM agent can do that job by itself: given target firing rates, drive simulators, checkers and optimizers until it finds parameters that are genuinely bistable.

I built a deterministic referee first, with no LLM inside: exact fixed points with Jacobian stability, cross-checked against a bit-reproducible RK4 simulator, plus an evaluation budget. Baselines were budget-matched and tested on held-out models (an E/I pair and the Wilson-Cowan 1972 model), with protocols written before each run.

The main finding is that the loss does most of the solving: CMA-ES on my expert loss beat every other setup, and roughly doubled the success rate of a naive loss on Wilson-Cowan. A local Qwen agent then solved 95% of the solvable episodes, close to the fixed pipeline’s 98%. But on tasks that provably have no solution, it claimed a false solution in 46 of 287 episodes, where the pipeline never did. A pre-registered test showed that adding a prompt sentence about scores did not fix it.

The codebase includes a leak-tested agent interface, full JSONL logging with episode replay, local LLM serving through llama.cpp, and CI with unit, golden-file and determinism tests.

Stack: Python, NumPy, SciPy, CMA-ES, Optuna (TPE), llama.cpp, Qwen, pytest, uv