Tirth Kanani

I am an engineer and researcher in London. Much of my work concerns how a language model can be given exactly the information a task demands, and nothing more.

That question led me to build code-review-graph, an open-source tool that maps a codebase so that a coding assistant reads only what a change requires. I am also building CrumbleUX, which watches a product in use and critiques it from six distinct points of view, much as a good design team would.

I work as an AI and automation engineer at Telana and hold the UK’s Global Talent visa. My MSc at the University of Birmingham asked whether a model running on one’s own machine could answer questions about private documents and show where each answer came from. Since then my interests have widened to graph theory, to the factors that shape a decision, and to machine learning models sound enough to generalise from one domain to another. People grasp unfamiliar situations quickly and from very little evidence, and I want to understand how a machine might reason as capably when matters are ambiguous. To that end I am building agents whose every decision can be examined afterwards, and testing them against the ARC-AGI-3 benchmark.