Tirth Kanani

I build infrastructure and tools that make frontier AI systems more reliable, interpretable, and efficient in the real world.

Most recently I created code-review-graph — a persistent, incremental code knowledge graph purpose-built as an affordance for LLM agents and coding agents. It uses Tree-sitter AST parsing across 12+ languages, multi-hop dependency reasoning via blast-radius analysis, and SHA-256 hashed incremental indexing. This lets it re-index a 2,900-file monorepo in under 2 seconds while delivering 6.8× average context compression (up to 49× on large monorepos). It is now shipped as an open-source Claude Code plugin with MCP server integration and has picked up over ten thousand stars on GitHub.

I am also the founder of CrumbleUX, where I ship a full-stack LLM product end-to-end (Python inference glue + TypeScript/Next.js/Node.js/PostgreSQL). I designed real-time Gemini Live streaming pipelines with bidirectional audio/video and structured prompting systems for reliable VLM critiques.

Previously I was a Graduate Research Assistant at the University of Birmingham's HCI & AI Lab, where I built multi-agent reliability evaluations and Graph Neural Networks that improved anomaly detection by 22%. My MSc thesis GraphMinds explored how knowledge graphs can make LLM systems more transparent and steerable. I graduated as valedictorian with distinction.

Based in London. Currently thinking about agentic systems, knowledge graphs, AI safety and alignment, context-efficient retrieval, and building AI that is both powerful and trustworthy.