Systems
Alongside papers, I build and run a small set of systems. They collect AI/ML papers from the official proceedings of 15 venues, process the full text and citations of about 185,000 papers, serve them on the web and to AI agents, and keep my curated paper lists current. Some started as research infrastructure, some as tools for my own day-to-day work. Together they run on four machines. These two pages show how.
Lab systems map
Every project, the machine it runs on, and the data that moves between them, from the venue websites to search results. In the interactive map you can click any box for details or trace one paper end to end.
The projects on the map:
- OpenPapers: collects PDFs and metadata from the official proceedings of 15 AI/ML venues.
- mustcite.com: web search over the collection, and the home of the MasterSet must-cite benchmark.
- paper-retrieval: search for AI agents over the full text and citation graph (below).
- paperlevelup: a Claude Code / Codex skill that organizes folders of papers and finds what is missing.
- Curated paper lists, refreshed from the collection: machine learning for graph drawing, neural algorithmic reasoning and combinatorial optimization, scientific literature retrieval, AI for vulnerability detection.
Inside paper-retrieval
paper-retrieval is a paper search service built for AI agents rather than people. An agent such as Claude Code asks a question over MCP. The service searches about 185,000 papers with dense and keyword retrieval, reranks the candidates, expands along the citation graph, and returns every result with the passage that supports it, so the agent can check its own answer. paperlevelup uses it to find papers.
The walkthrough follows one real search step by step, shows how the index is built, and lists the tools an agent can call.