I'm Thomas van Dongen. I build ML systems for embeddings, retrieval and language models. I'm head of AI engineering at Springer Nature and co-founder of Minish, an open-source ML lab where we build small, fast models and packages for embeddings and search.
- semble: code search for agents that uses ~99% fewer tokens than grep+read
- model2vec: state-of-the-art static embedding models that are 500x faster than transformers
- semhash: semantic deduplication, outlier filtering and representative sampling
- pyversity: diversify search and retrieval results, with NumPy as the only dependency
- vicinity: lightweight nearest neighbor search with one interface for many backends
- model2vec-rs: Model2Vec inference in Rust
- pi-extensions: minimal extensions for the Pi coding agent