Gagan DevagiriGagan Devagiri

The current safety problem in Agentic AI rarely stems from abstract misalignment. It is a team deploying an agent without mapping the surface area it is exposed to at runtime. I’m building Pome to close that gap, giving agents a safety infra grounded in real conditions.

Studied Artificial Intelligence at the University of Edinburgh. Interned as a Software Engineer at Google and as an ML Engineer at TikTok, then joined Thought Machine, a fintech unicorn in London as a new grad. Left to start my own company and build from 0 to 1.

I also like to teach. I TA’d Edinburgh’s Functional Programming and Robotics courses for multiple semesters. Outside tech, I played competitive League of Legends — peaking Top 100 on the EUW ladder against LEC players. These days I’m more into building Agent Harnesses and fine-tuning edge models.

Pi Coding AgentBenchmarked, context-efficient tools for the Pi Agent Harness (10k+ monthly downloads)
Gameboy Emulator - WebMCP ArenaA Game Boy emulator in the browser with WebMCP access for AI agents to play Pokemon Red.
Qwen 10Post-trained Qwen 3.8-27B with RL to improve its performance for LLM-as-a-judge.
Voice-Agent Personal ShopperVoice-first interface for all shopping apps. Built at Sarvam Epoch (Top 5/300+ projects)
Research Search EngineSearch millions of papers in a fraction of a second.
PolymorphA local log compressor that beats Microsoft's LLMLingua-2 by 3x while keeping the details LLMs need to debug.

I also write about what I'm building and learning. Check out my notes.