$ whoami --verbose
Arsh Chawla // machine learning researcher
I build systems that reason: imitation learning for robot manipulation, geometry-grounded benchmarks for LLM attention, and research tools that can't hallucinate their sources. I like working close to the metal too: GPU programming, hardware, and getting low with Rust.
01 // publications
02 // selected work
BLOG.001Can Algorithmic Trading Signals Catch an LLM Hallucinating?
An LLM's per-token uncertainty is a time series, and quant finance has spent decades building interpretable signals for exactly that shape of problem. I backtested trading strategies against 788 judged answers to see which ones catch the model making things up.
BLOG.002OpenMyst: Anchored LLM Information Retrieval Tool for Research
OpenMyst is an anchored-research tool where every claim is tied to a verbatim line in a real source, shipped as an MCP connector and a full multi-agent desktop app. Built to solve a pain point, not to make money.
BLOG.003ZelusBench: Measuring LLM Attention with Geometry
How I built a deterministic, geometry-grounded attention benchmark for the Google DeepMind Measuring Progress Toward AGI hackathon.
BLOG.004RiddleForge: iOS Puzzle Game
The Riddle Forge is an LLM based brainteaser app, out now on the iOS App Store!
RES.001LSTM-based Cyclone Forecasting
A lightweight LSTM for forecasting tropical cyclone tracks that outperformed heavier deep-learning baselines.