I build RAG over 100k documents, LLM agents,
and Deep Learning trained on supercomputers.
4 years at CNRS training models for researchers in weather, astrophysics, ecology and nanomaterials. Multi-GPU on Jean Zay, libraries still running in prod at Météo France, 2 peer-reviewed publications.
Now at Berger-Levrault, I build Athena: business users ask questions about ~100k technical documents, RAG and MCP agents find the answer and cite their sources.
3 packages on PyPI. When I'm not coding, I read sci-fi and hike the Pyrenees. Based in Toulouse, open to Paris / Montreal.
Berger-Levrault
AI R&D team · 12 people
Athena is Berger-Levrault's agentic platform: tool-first agent, connected to documents and business APIs for all clients (local government, industry, maintenance).
In production, ~70 pilot users. Langfuse observability, client workshops, iterating on field feedback.
4 key contributions:
Athena · tool-first agentic architecture
Context: build the agentic foundation for all BL clients (local government, industry, maintenance).
Role: agentic architecture rebuild with the team (LangGraph): from a router (one agent per task) to a single tool-first agent — context management, skills, sub-agent orchestration, source-grounded answers.
Result: platform in production, source-grounded and actionable answers across multiple business domains.
Content-extractor · OCR/PDF/DOCX
Context: initial service was a basic image extractor and hard to extend.
Role: full OCR/images/PDF/DOCX rebuild, async batching (Celery + Mistral batch API), factory/registry architecture to add extractors/chunkers quickly.
Result: -50% extraction costs.
Airflow pipelines
Context: document ingestion per client account.
Role: evolved the ingestion pipelines with the team (PDFs, manuals, work orders, machine docs) and added RAG enrichment steps: data augmentation, chunking, embeddings, question/keyword generation for indexing.
Result: 5 operational DAGs (1 per client).
MCP Builder
Context: rapidly connect BL's internal APIs to agents.
Role: an LLM agent analyzes OpenAPI specs to pre-process endpoints (grouping, masking, description generation) and generate MCP servers via FastMCP, then human-in-the-loop review to validate routes and business context.
Result: ~100 internal APIs turned into MCP servers, deployed.
CNRS · National AI Research Program (PNRIA)
Part of the PNRIA network, I worked with research teams across France on their AI problems.
Always 2 projects running in parallel (6-12 months each), delivered to partners like Météo France, CNES, CEA, INEE.
Training and fine-tuning on Jean Zay (multi-GPU DDP, up to 8 GPUs, Slurm). PyTorch profiling.
Selected proof points:
GENS / MetScore · Météo France
Context: evaluate weather models in production, multi-GPU optimization and fine-tuning of a diffusion model.
Role: designed MetScore (YAML configuration + Python library) and trained a diffusion model (DDPM) in PyTorch for weather-image generation on Jean Zay.
Result: library still in production + diffusion POC at -20% compute with equivalent quality.
Co-author on AMS 2025 paper
DeepFaune · CNRS/INEE
Context: wildlife recognition on camera traps (1.5M images).
Role: fine-tuning YOLOv8 on custom dataset (1.5M images, 24 classes), multi-GPU training, class-imbalance handling, precision/speed optimization for CPU deployment.
Result: 93% on 24 species, 3× faster.
Peer-reviewed publication
Other contributions: AUTOFILL (CEA, PairVAE, MAE 0.98), BIGSF (CNES, tech lead and architecture refactor), Introduction to LLMs training (3h, ~25 CNRS PhD students/researchers).
Agileo Automation (apprenticeship)
Supervision and control framework for robotic machinery in semiconductor manufacturing. Object-oriented architecture, C#, HMI. Team of 5 engineers, Agile/Scrum.
M.Sc. · Artificial Intelligence & Pattern Recognition (IARF)
Université Paul Sabatier Toulouse III · IRIT
Specialization in Deep Learning, Computer Vision, NLP. IRIT lab (Toulouse Computer Science Research Institute).
B.Sc. Computer Science
Université Paul Sabatier Toulouse III
Languages
LLM mock server to test retries, fallbacks and rate limiting, zero tokens spent
Denoising diffusion probabilistic model for weather imagery · −20% resource usage
Python utility to generate unique file paths: incremental suffixes, UUIDs, timestamps
End-to-end ML pipeline deployed on AWS · Computer Vision + REST API
"Enriching Operational High-Resolution Ensemble Forecasts with StyleGAN-2"
C. Brochet, G. Moldovan, J. Rabault, C. Regan, L. Raynaud
Artificial Intelligence for the Earth Systems (AIES), vol. 4, no. 1 · DOI: 10.1175/AIES-D-24-0058.1
"The DeepFaune initiative: a collaborative effort towards the automatic identification of European fauna in camera trap images"
Julien Rabault et al.
Google Scholar
Agentic AI & RAG
Deep Learning & Vision
MLOps & Infrastructure
Engineering & Frameworks
Contact
Available for new opportunities. Toulouse, Paris or Montreal. Remote OK.
Domains: GenAI · RAG · Multi-agent Systems · MLOps · Computer Vision
Download my CVJulien Rabault · Toulouse, France ·