PhD in Computer Science · Data Scientist & AI Engineer · NLP Specialist
Aix-en-Provence, France · Open to new opportunities
Data Scientist and AI Engineer with a PhD in Computer Science (Aix-Marseille Université, 2022), specialising in NLP, LLMs, RAG and generative AI. I combine published ML research with hands-on product engineering.
- 🎓 PhD in Computer Science — Aix-Marseille Université, LIS / IrAsia — ERC Advanced Grant ENP-China (n° 788476)
- 📄 7 peer-reviewed publications — LREC-COLING, NLP4DH, TALN, PACLIC, JDMDH, JHNR
- 🌍 Presented research in 7 countries (France, USA, China, Italy, Japan, India, Vietnam)
- 🔭 Currently building a Data & AI SaaS platform — RAG pipeline, text-to-SQL, multi-tenant architecture
Languages
ML / AI
LLM & Generative AI
litellm · LangGraph · Langfuse · RAG / GraphRAG / RAPTOR · Hybrid Search (BM25 + pgvector) · Reranking · Text-to-SQL · Prompt Engineering · VLM
Full-Stack & Infrastructure
| Project | Description | |
|---|---|---|
| HistText | Full-stack platform for large-scale analysis of historical Chinese texts (billions of tokens). Rust backend, React UI, Apache Solr, multilingual NER pipeline, R package on CRAN. Deployed for the international digital humanities community. | 🌐 Live |
| EventExtractionPapers | Curated and actively maintained list of NLP papers, datasets and models for event extraction. Widely used reference in the research community. | ⭐ 581 |
A daily AI/ML · DevOps · Cloud digest I generate automatically from my own tech-watch — no human in the loop.
- Thinking Machines has released Inkling 2 Small, pretrained from scratch in under 3 days on fewer than 12T tokens in native NVFP4, with a single joint RL run, and it is reported as competitive with, and often beating, same-size Chinese models.
- GitHub's usage metrics API now reports, per repository, how many Copilot code review comments developers acted on, with a computed addressed rate (addressed comments divided by evaluated comments) and breakdowns by severity, review tier, comment type and customization.
- The same API now flags users whose client telemetry is missing or whose IDE or Copilot extension versions are below the usage metrics minimums, with daily affected-user counts at enterprise and organization level.
- Docker Sandboxes are presented as a distinct technology from containers: dedicated microVMs with their own Linux kernel, filesystem, network and Docker Engine, built for autonomous coding agents, whereas containers share the host kernel.
- Pinecone's webinar recap argues that AI agents need more than vector similarity: a semantically relevant result can still give a wrong answer when, for example, a one-day promotion is counted as demand, so the decision logic belongs outside the search step.
➡️ Full digest & archive · updated twice a day, no human in the loop.