Vyomakesh Dundigalla

I build evaluation infrastructure, open-source RL environments, and GPU-level performance tooling for AI systems.

Now

Working on ARC-AGI-3, building open-source RL environments.

Selected work

Made Zerobrew downloads 3.6× faster (361 s → 99 s) with racing cancellation and HTTP/2 tuning, merged to main and shipped to every user.

Built six reproducible evaluation environments (safety, web agents, tool use/MCP, and a game suite) for Prime Intellect's Community Environments, separating rollout, verifier, and scoring paths so failures are easy to attribute.

Reached 12.804 µs on an NVIDIA B200 in GPU Mode's nvfp4 block-scaled GEMM hackathon, writing CuTeDSL kernels aimed at speed-of-light performance.

Built OpenGraphs at Build India Hack (Anthropic × Replit × Lightspeed): a local-first, TUI-native experiment tracker for AI runs over SSH, with live graphs, logs, and run comparison on a Rust-first stack.

Research

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