|| ఓం ||

Vyomakesh Dundigalla

Over the past couple of months I have contributed to Prime Intellect's Environments Hub, now Community Environments, building eval setups that are easier to run, debug, and trust.

I also worked on Zerobrew and helped push the download path from around 6 minutes to roughly 1 minute. From there I went deeper on GPU kernels with PTX, CUTLASS, and CuTeDSL, and took part in the NVIDIA FP4 GEMM optimization hackathon where we pushed kernels close to speed-of-light performance.

One project I really enjoyed was OpenGraphs, which we built during Build India Hack by Devfolio (Anthropic x Replit x Lightspeed), a local-first real-time TUI for AI training that helps developers spot issues early and cut wasted GPU hours.

Vyomakesh Dundigalla profile photo

Projects

OpenGraphs is a local-first, TUI-native experiment tracker for AI runs over SSH, with live graphs, logs, run comparison/filtering, and a Rust-first stack (ratatui + backend daemon + optional Python agent layer).

Proposals

VGAC: Verifier-Grounded Agreement-Calibrated Reinforcement Learning for Long-Horizon Agents explores verifier-grounded rollout agreement, entropy-aware calibration, and reward shaping to reduce silent overconfident failures in long-horizon RL agents.