hugging face · crates.io · pypi

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 an industrial AI lab. Today I work on LLM agents and how to evaluate them, NLP and retrieval over biomedical literature, and small language models fine-tuned for one task and cheap enough to serve.

At my day job I work on a private NLP and LLM platform for pharma, including its model server and evaluation harness, so what I can share is below.

llms       →  agents, RAG, LLM-as-judge evaluation, fine-tuning (SFT, DPO, GRPO), synthetic data
nlp        →  classification, NER, embeddings and retrieval, reranking, question answering
vision     →  uncertainty and active learning on images, image quality, vision-language models
ml         →  model merging, distillation, gradient boosting, Bayesian optimization
serving    →  inference optimization, quantization, dynamic batching, model servers

Then You Are · play it · model
A word game you play against friends or a fine-tuned 0.8B model, while another model judges every move against the game's rubric. On game variants it never saw, 68.5% of the fine-tuned model's moves hold up against the judge, up from 19% before training.

SIREN · report · models
Retrieve-then-rerank for screening biomedical literature in systematic reviews. It uses a bi-encoder and a cross-encoder trained on synthetic query-document pairs built from PubMed, and the dataset and both models are on Hugging Face.

ModernBERT SQuAD 2.0 · base · large
ModernBERT-base and ModernBERT-large fine-tuned for extractive question answering on SQuAD 2.0. The base model has been downloaded more than 26,000 times on Hugging Face.

abbreviation-extractor · crates.io · PyPI
Finds abbreviation-definition pairs in biomedical text, with an improved Schwartz-Hearst algorithm in Rust and Python bindings.

blitztext · crates.io · PyPI
Fast keyword extraction and replacement in strings, based on FlashText and Aho-Corasick, in Rust with Python bindings.

autoregressive-text-classification
Fine-tunes a 270M generative model to classify text and compares it with a DeBERTa encoder. It reaches 98.6% of the encoder's F1.

_{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}