AI / ML scientist & engineer
Yannick Detrois
I work on compact, high-performance language models — synthetic data, from-scratch pretraining, distributed training, and getting small, specialised models into production. AI/ML Scientist at Pleias and an MSc student at EPFL.
German & French (native) · English C2 · Spanish (basic)
experience
AI Researcher · Freelance · Tw0zer0s
- ▪Independent consultant on the data side of pre- and post-training LLMs — dataset curation, OCR / NLP components, and training of small, specialised models.
- ▪Partnering with tech companies, research labs, and organisations in regulated industries.
AI / ML Scientist · Pleias
- ▪Co-built SYNTH — the first open generalist synthetic pretraining dataset for small reasoning models (75B tokens, 8 languages) — and pretrained the Baguettotron models (56M–600M dense, 13B/1B MoE). All open on Hugging Face; NeurIPS 2026 in review.
- ▪Co-author, ACL 2026 Findings: a 600M synthetic-only reasoning model matching proprietary LLMs >1000× its size on customer-distress classification, now in production.
- ▪Co-released Nemotron-Personas-France with NVIDIA: 1M census-grounded synthetic personas via a probabilistic graphical model.
- ▪Distributed training of multimodal models, large-scale inference, and public talks (ACL 2026, BNF, VivaTech).
Data Analyst & ML Consultant · Bacardi-Martini
- ▪Cut statistical sales-forecast error by 30%+, deployed globally across ~100 markets for all products.
- ▪Led Python data processing, custom error-metric design, and backtesting + live validation of time-series models in SAP IBP.
Research Computing Specialist — Process Modelling & HPC · Micari Group, EPFL
- ▪Optimised a carbon-capture adsorption model with genetic algorithms (pymoo NSGA-II) and built an HPC pipeline for scalable CPU/GPU workflows.
- ▪Led infrastructure setup and team enablement on a return internship. Paper under review at Nature Sustainability (ChemRxiv preprint).
papers
Model in Distress: Sentiment Analysis on French Synthetic Social Media
Pierre-Carl Langlais, Pavel Chizhov, Yannick Detrois, Carlos Rosas Hinostroza, Ivan P. Yamshchikov, Bastien Perroy
SYNTH: training reasoning models with generalist synthetic environments
Pleias team (incl. Yannick Detrois)
Optimization of Hybrid Direct Air Capture unlocks economically feasible process configurations
Moein Fazli Mamaghani, Daniele Micucci, Stefania Bempeli, Shivang Sachar, Yannick Detrois, Himan Dev Singh, Wendy L. Queen, Kumar V. Agrawal, Marina Micari
selected work
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.
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.
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.
education
off the clock
coming soonwork in progress
A corner for everything outside the models
Mostly mountains and adventures — paragliding, alpinism, climbing, backcountry skiing, trail running, and long days on the bike — plus teaching, competition maths, and a saxophone. Full stories, photos, and routes are on the way.
Get in touch
Let's talk
Always happy to discuss small language models, synthetic data, or a potential collaboration. Reach me by email, LinkedIn, or X.