Most of my production work lives in private client repositories, so it can't be shared. What I can share is here: runnable reference code for my blueprints, and the open-source projects I think deserve more attention.
The working version of every blueprint on this site: chunking strategies you can compare on your own documents, long-document retrieval with a heading tree, tiered OCR routing and tiered agent memory. Dependency-free core, tested in CI.
Retrieval evaluation with a regression gate that uses a paired bootstrap, so CI fails on real drops and ignores noise, plus an LLM judge you calibrate against human labels before trusting it.
A drag-and-drop workflow platform in the spirit of n8n: a React Flow canvas, a backend engine that executes saved workflows in real time, trigger, action and if/else nodes, and an AI summary step.
Also on my GitHub
Research code I keep on my GitHub because I use and recommend it. Each is the original authors' work, with their licence intact.
Chunk embeddings that carry the whole document's context, with no extra model calls. The cleanest fix I know for chunks that lose meaning on their own.
Recursive summary trees for answering questions about a whole document. The idea behind my long-document design.
Retrieves long units instead of short passages, so long-context models see coherent evidence. A useful counterweight to over-chunking.
Skip retrieval entirely for small, static knowledge bases by preloading and caching them. Worth knowing before building a RAG stack you don't need.
RAG evaluation that works without golden answers. Handy when you have real traffic but no labelled set yet.
Training and running late-interaction (ColBERT-style) retrievers. The most active home for multi-vector retrieval.
Agent memory organised as linked, evolving notes. A thoughtful alternative to flat vector memory.
The benchmark I'd use for agent memory: updated facts, reasoning across sessions and over time, and knowing when to say 'I don't know'.
Projects I rely on
Well-built projects I build on. Each belongs to its authors.
Temporal knowledge graphs where changed facts are invalidated, not deleted. The model for how I handle facts that change.
MIT-licensed document conversion that keeps headings, tables and reading order. Structure-aware chunking starts here.
An open OCR model and benchmark that holds up on hard pages, with published cost per million pages.
Open embedding and reranker models that run on one GPU. The reranker is the cheapest big win in most RAG stacks.
Contact
I'm open to AI engineering, architecture and training work. Tell me what you're building and what the constraints are — that's usually enough to start.