Thomas van Dongen
Head of AI Engineering at Springer Nature · Co-founder of Minish
I build production ML systems for embeddings, retrieval, and language models. At Springer Nature I lead a team building AI-driven tools for scientific publishing: research discovery, compliance screening, and recommender systems that run at scale.
At Minish, the open-source ML lab I co-founded with Stéphan Tulkens, we build small, fast packages and models for embeddings and search. Our models are hundreds of times faster than transformers at a fraction of the size.
Open source
All projects →- Semble Code search for agents that uses ~99% fewer tokens than grep+read.6.2k 549.8k
- Model2Vec Distil any sentence transformer into a tiny, fast static embedding model.2.2k 6.9M
- SemHash Semantic deduplication and dataset filtering across text, images, and audio.972 873.3k
- Pyversity Diversify search and retrieval results to reduce redundancy and improve coverage.501 537.6k
- Potion Tiny state-of-the-art static embedding models for English, multilingual, code, and retrieval.
Writing
All posts →- Agent-First vs Agent-Second EngineeringWho writes the first code matters more than who writes the rest.Sep 2026
- Model2Vec: Distill a Small Fast Model from any Sentence TransformerDistill small, fast static models from any Sentence Transformer without needing a dataset.Oct 2024
- Demystifying Efficient Self-AttentionA practical overview of efficient attention mechanisms that tackle the quadratic scaling problem.Nov 2022