Thomas van Dongen
Thomas van Dongen is an individual machine-learning engineer, open-source developer, and technical writer—not an organization. He builds production ML systems for embeddings, retrieval, and language models, and co-founded Minish, an open-source ML lab focused on small, fast models and packages for embeddings and search.
At a Glance
- ML and NLP developers
- AI platform and search teams
- Developers building RAG, classification, clustering, or recommendation systems
- Data and ML teams cleaning and curating multimodal datasets
- +2 more
AI Tools by Thomas van Dongen
(1)
pi-extensions by Pringled
Pi Coding Agent Extensions Collection
Discussions
No discussions yet
Be the first to start a discussion about Thomas van Dongen
Latest News
Published 'Agent-First vs Agent-Second Engineering,' an essay on who writes the first code in agentic engineering.
Semble documented/released as fast and accurate code search for agents, using about 99% fewer tokens than grep plus read.
pi-extensions repository reached a latest commit adding click-to-expand folded tool rows; the project lists nine installable Pi extensions.
SemHash was published as a Minish project for fast multimodal semantic deduplication and filtering.
Products & Services
A technique and Python package that distills Sentence Transformers into tiny, fast static embedding models using vocabulary representations, PCA, and Zipf weighting; supports distillation and inference and integrates with Sentence Transformers, LangChain, and LlamaIndex.
A lightweight multimodal semantic-deduplication and dataset-filtering toolkit for text and images, including deduplication, outlier filtering, and representative sampling.
Code search for AI agents that combines code-aware chunking, Model2Vec embeddings, BM25, reciprocal-rank fusion, and reranking; it can run locally on CPU and integrates through MCP, CLI instructions, or a sub-agent.
A NumPy-only search/retrieval diversification package intended to reduce redundancy and improve result coverage.
Market Position
Thomas's open-source work is positioned around small, fast, local, CPU-friendly alternatives to transformer-heavy embedding and search stacks. Model2Vec emphasizes speed and compactness versus Sentence Transformers; SemHash and Vicinity target lightweight dataset and nearest-neighbor workflows; Semble combines Model2Vec with BM25 for agent code search and claims roughly 99% fewer tokens than grep-plus-read. Relevant comparison points include Sentence Transformers, GloVe/word2vec-style embeddings, transformer-based code search, BM25/grep-based retrieval, and vector-search backends.
Leadership
Executive Team
Thomas van Dongen
Head of AI Engineering at Springer Nature; co-founder of Minish
Studied Artificial Intelligence at the University of Groningen; joined Slimmer AI and led its recommenders team in the scientific domain; after Springer Nature acquired Slimmer AI's science division in 2023, he became Head of AI Engineering. He also maintains the Pringled open-source projects and co-founded Minish with Stéphan Tulkens.
Stéphan Tulkens
Co-founder of Minish
Thomas's co-founder on Minish and co-author of the Model2Vec and SemHash projects.
Founding Story
Thomas's personal work grew from his AI studies at the University of Groningen, followed by work at Slimmer AI on scientific-domain machine learning and recommender systems. He and Stéphan Tulkens later co-founded Minish to build small, fast, hardware- and eco-friendly embedding/search models and packages; his personal site says the models are intended to be dramatically faster and smaller than transformer alternatives.
Business Model
Revenue Model
The publicly described work is open-source software and models, generally distributed under permissive licenses or through public package/model registries. No paid subscription, API pricing, or commercial revenue model is stated on the reviewed personal and project sources.
Target Markets
- ML and NLP developers
- AI platform and search teams
- Developers building RAG, classification, clustering, or recommendation systems
- Data and ML teams cleaning and curating multimodal datasets
- AI coding-agent and developer-tool users
- Scientific publishing and research-discovery organizations
- Search and retrieval
- Semantic search and retrieval-augmented generation
- Text classification and clustering
- Dataset deduplication, filtering, and sampling
- Code search for AI coding agents
- Scientific-literature discovery, compliance screening, and recommender systems
- Springer Nature