# Tirth Kanani

> AI engineer and researcher building context systems, developer tools and auditable agents in London.

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

I created [code-review-graph](https://github.com/tirth8205/code-review-graph) and I am building [CrumbleUX](https://crumbleux.com). Both turn messy signals into useful decisions.

By day I am an AI and automation engineer at Telana in London, endorsed by the UK as Global Talent. My MSc research at Birmingham explored traceable local language models; now I am testing auditable agents against ARC-AGI-3.

## Recently

- Oct 2026: [Built Gateway Check at the Future States Hackathon, No. 10](https://github.com/tirth8205/No10_Hackathon). One of 35 engineers chosen from more than 1,500 applicants, run by 10DS and the Cabinet Office. Gateway Check reads domestic abuse evidence for legal aid against Schedule 1 before it is submitted, and every verdict rests on a quote found word for word in the letter.
- Sep 2026: [Demoed code-review-graph at Claude Founder House](https://www.linkedin.com/feed/update/urn:li:activity:7508887893927469056/). Live on stage at Anthropic’s day for founders and builders in London. Median 8.9× fewer tokens across 18 questions on six real repositories.
- May 2026: [Won the Build with AI Hack Night at Google I/O London](https://www.linkedin.com/feed/update/urn:li:activity:7463224958392131584/). Solo, with CrumbleUX built in one night during my first week at Telana. Hosted by GDG London.

## Selected work

- [code-review-graph](https://tirthkanani.com/writing/code-review-graph.md) (2026): Local codebase context for coding agents. 32,222 GitHub stars · up to 49× fewer tokens.
- [CrumbleUX](https://tirthkanani.com/writing/crumbleux.md) (2026): An AI UX team that watches products being used. Founder · product, inference and infrastructure.
- [Harmonic Flow Models for ARC-AGI-3](https://tirthkanani.com/writing/hfm-arc-agi-3.md) (2026): Auditable exploration for interactive abstract reasoning. Presented at Pre-ICML @ London 2026.
- [Jailbreak-Eval](https://tirthkanani.com/writing/jailbreak-eval.md) (2025): Adversarial testing for language models. Five attack generators · reproducible pilot baselines.

## Writing

- [The Coming Intelligence Import Bill](https://tirthkanani.com/writing/intelligence-import-bill.md) (2026): Why access to frontier AI is becoming a strategic trade relationship, and what countries should build before the API gateway acquires a toll.
- [From rule-based algorithms to agentic AI](https://tirthkanani.com/writing/agentic-ai.md) (2025): How software went from fixed rules to language models to agents that plan and act, and what that shift asks of the people building it.
- [How I got here](https://tirthkanani.com/writing/about.md) (2025): From taking apart radios in Ahmedabad to building AI systems in London, via a brain-computer interface and a hackathon win.

## Earlier projects

- [EEGSpeech](https://tirthkanani.com/writing/eegspeech.md) (2025): A brain-computer interface that decodes imagined speech from EEG signals using a CNN-LSTM model, achieving over 92% accuracy.
- [GraphMinds](https://tirthkanani.com/writing/graphminds.md) (2025): An MSc thesis project that combines knowledge graphs with local LLMs for private, transparent analysis of unstructured documents.
- [ResearchWebGraph](https://tirthkanani.com/writing/research-web-graph.md) (2025): An open-source tool for exploring academic papers through knowledge graph visualisation and AI-grounded question answering.

## Elsewhere

- [X](https://x.com/tirth_8205)
- [GitHub](https://github.com/tirth8205)
- [LinkedIn](https://linkedin.com/in/tirthkanani)
- [Email](mailto:tirthkanani18@gmail.com)

## For agents

This page is the plain-text twin of https://tirthkanani.com. Every article has one too: add .md to its address. A short index for language models lives at https://tirthkanani.com/llms.txt.