selected & applied work
projects
Research and applied work across synthetic data, compact language models, distributed training, and production ML.
How small can a verifier be?
How verifier reliability scales with model size across three tasks — down to 0.07M-param models pretrained from scratch, benchmarked against a frontier judge. Interactive charts + full methodology.
SYNTH + Baguettotron
The first open generalist synthetic pretraining dataset for small reasoning models (75B tokens, 8 languages), and a family of models pretrained on it at SoTA data efficiency. NeurIPS 2026 in review.
Nemotron-Personas (with NVIDIA)
Synthetic national personas grounded in census data via a probabilistic graphical model — open-sourced with NVIDIA to unlock AI where real data is sensitive. Now covering France and Belgium.
RATP distress detection (ACL 2026)
Model in Distress: a 600M synthetic-only reasoning model that matches proprietary LLMs >1000× its size on customer-distress classification — the first industrial deployment of our synthetic-data work, now in production.
GPU MODE — 2nd place
Pretrained language models in 10 minutes on 32 GPUs (FP8, FlashAttention-4), exploring 100M–1B scaling under fixed compute at the PyTorch / GPU MODE / Neon Noir hackathon.
Bacardi global forecasting
Cut statistical sales-forecast error by 30%+ and deployed the system globally across ~100 markets — custom error metrics, backtesting, and live validation in SAP IBP.
Hybrid Direct Air Capture (HPC)
Optimised a carbon-capture adsorption model with genetic algorithms (NSGA-II) on an HPC pipeline, unlocking economically feasible configurations. Under review at Nature Sustainability.
Soft compression of long-context prompts
EPFL DLAB semester project: training an LLM to compress long documents into latent space for cheaper caching and context-length extension.