about
this bio is 107 tokensModernBERT tokenizer
I'm an AI engineer in Paris, and I've been building ML systems since 2020, from a Japanese conversational model in Hanoi to research on uncertainty and active learning for image models at the SLB AI Lab.
At ArcaScience I work across an AI benefit-risk platform for pharma: LLM agents and how we evaluate them, NLP pipelines over biomedical literature, and the model serving behind them. I also develop small language models, fine-tuned for a single task and cheap enough to serve in production.
what I work on
- LLM agentsRAG, tool use, fine-tuning (SFT, DPO, GRPO), synthetic data, model merging, distillation
- NLPclassification, NER, question answering, biomedical text
- retrievalembeddings, dense retrieval, reranking, search over scientific literature
- computer visionuncertainty and active learning on images, image quality, vision-language models
- evaluationLLM-as-judge, eval harnesses, ablations, uncertainty estimation
- model servinginference optimization, quantization, dynamic batching, model servers
where I've worked
- 2023 to nowArcaScience, ParisAI engineer on LLM agents, evaluation, biomedical NLP and model serving
- 2023SLB AI Lab, ClamartUncertainty estimation and active learning, my master's thesis
- 2022ArcaScience, ParisData augmentation and robustness for biomedical NLP models
- 2020 to 2021Aimesoft, HanoiA Japanese conversational model, and software for AI avatar and receptionist products
- 2017 to 2023USTH Hanoi, ESIEE ParisBSc in ICT, then an engineering degree in AI and cybersecurity
papers and theses
- AI-Powered Scoring System for Detecting Drug-Drug Interaction Insights from Scientific LiteratureInnovative Insights in Digital Health, 2026
- Training an AI model for publication type classification: a case study on alopecia areataJournal of Medical Artificial Intelligence, 2026
- Uncertainty Estimation and Active LearningMaster's thesis, SLB AI Lab and ESIEE Paris, 2023
- Developing pre-trained models and other NLP models for chatbot systemsBachelor's thesis, USTH Hanoi, 2020
side projects
/
- Then You Are2026A word game played against a fine-tuned 0.8B modelthenyouare.com
- SIREN2026Retrieval models for screening biomedical literaturegithub.com
- ModernBERT SQuAD 2.02025Extractive question answering, 26k+ downloadshuggingface.co
- autoregressive-text-classification2025A 270M generative model as a text classifiergithub.com
- abbreviation-extractor2024Abbreviations in biomedical text, in Rustgithub.com
- blitztext2024Fast keyword extraction, in Rustgithub.com
Then You AreSIRENModernBERT SQuAD 2.0autoregressive-text-classificationabbreviation-extractorblitztextDrug-drug interaction scoringPublication type classifierMaster's thesisBachelor's thesis
toolkit
- LLMs and agents
- vLLM, llama.cpp, TensorRT-LLM, LangChain, DSPy, LiteLLM, Langfuse, MCP, OpenAI and Anthropic APIs, Vercel AI SDK
- Training
- PyTorch, Transformers, TRL, PEFT, Unsloth, Accelerate, bitsandbytes, mergekit, TensorFlow, Keras
- Classical ML
- scikit-learn, XGBoost, LightGBM, CatBoost, Optuna, UMAP, HDBSCAN, W&B
- NLP and search
- sentence-transformers, spaCy, medspaCy, GLiNER, Flair, NLTK, gensim, KeyBERT, PyLate, Elasticsearch, Solr, Spark NLP
- Vision
- OpenCV, timm, DINO, YOLO, Albumentations, scikit-image
- Uncertainty
- GPyTorch, BoTorch, Bayesian-Torch
- Evaluation
- RAGAS, MTEB, seqeval, HF Evaluate, matplotlib, seaborn
- Serving and MLOps
- ONNX Runtime, Optimum, CTranslate2, llm-compressor, FlashAttention, Liger Kernel, TEI, FastAPI, Gunicorn, Docker, Kubernetes, Helm, Traefik, Airflow, MLflow, OpenTelemetry, Gradio
- Data
- pandas, Polars, Ray, Pydantic, Playwright, PyMuPDF, GROBID, Label Studio
- Cloud and CI
- GCP, Cloud Run, Azure, AWS, Modal, SkyPilot, dstack, GitHub Actions, GitLab CI
- Languages and web
- Python, Rust, TypeScript, SQL, PostgreSQL, React, Next.js, Node.js