Arsh Chawla

A research collaborator that physically cannot invent a citation. Every claim it writes is anchored to a verbatim line in a source you gave it—and if there’s no anchor on disk, the claim doesn’t get written. Shipped as an MCP connector and a full multi-agent app, given away because it solves a real pain point.

For the Google DeepMind AGI hackathon I took on the attention pathway—a genuinely hard thing to measure. My answer was to make relevance a mathematical fact instead of a judgement call: ZelusBench grounds attention tasks in 3D geometry, and the results reveal distinct cognitive profiles across frontier models.

Ever since I was young, I’ve always loved solving riddles. But for me, the real magic of a riddle wasn’t just in reading a question and reasoning out an answer, it was in the interaction. This led me to developing RiddleForge, an LLM-engine based riddle game!

This research project explores LSTM RNNs for cyclone forecasting, comparing their accuracy and efficiency to state-of-the-art NeuralGCM and HAFS models.