Tirth Kanani

I build tools that give AI systems the right context: less of it, better chosen.

Most recently I built code-review-graph, a local knowledge graph that helps coding agents work in large repositories without reading the whole codebase.

It parses source code with Tree-sitter, maps dependencies, and hands the agent only the files and symbols a change touches. On big repositories that cuts token usage by up to 49 times. The open-source Claude Code plugin has picked up 27,891 stars on GitHub.

I also run CrumbleUX, an AI product that watches your screen and critiques the UX aloud, like a designer sitting beside you. I build and operate it end-to-end: Python for inference, TypeScript and Next.js for the product. The difficult part is not streaming live audio and video to a model. It is getting critiques back that hold up.

Previously I was a research assistant in the University of Birmingham's HCI & AI Lab. My MSc thesis, GraphMinds, used knowledge graphs to help a local language model show where its answers come from. I graduated as valedictorian.

Based in London. Interested in agents that use less context, show their sources and fail more safely.