CS at BITS Pilani. I work on AI infrastructure and spend most of my time trying to break it.
74 PRs merged upstream, 56 of them AI work, across 12 organisations including Microsoft, PyTorch, NVIDIA, Hugging Face and CNCF.
- Nine of my patches are in GCC master, carried up from the gccrs tree by the project's syncs to gcc-patches with authorship kept.
- 40 of those 74 fix bugs I found and reported myself, not tickets off a board. One was in my own merged code.
- I found nondeterminism in vLLM's deterministic mode (vllm#51187): repeats match only if every token is reduced at the same RMSNorm block width.
- Sole author of a COLM 2026 workshop paper. When Self-Consistency Backfires: majority voting lowers accuracy on 56–66% of GPQA Diamond problems.
| Project | Merged |
|---|---|
| microsoft/PyRIT | 37 · the CodeAttack converter, plus fixes to TAP/PAIR, scale scorers and dataset loaders |
| gccrs · GCC Rust frontend | 11 · match arm guards, left-to-right argument evaluation, a dead-code fix now in GCC master |
| PyTorch · executorch, torchtitan | 8 · Inspector crashes on non-finite output, wandb tag splitting, scheduler range checks |
| jaeger-ui · CNCF | 4 · GenAI span classification, media rendering, RFC wire-format corrections |
| pwndbg · GDB exploit-dev frontend | 3 · a pointer-scan crash, its help text, negative-value searches |
| OpenEnv · was Hugging Face | 2 · SSRF-safe URL parsing, Space subdomain resolution |
| meta-pytorch/tritonbench | 2 · child runs failing unless started one exact way, A/B metrics crash |
| NVIDIA/garak | 1 · fixed Bedrock scans that failed for every Claude 4.x user |
| Also | llm-compressor ×2, AI Village ×2, openkruise/agents, openyurtio/raven |
| Lockstep | Verifies an inference engine returns the same tokens whatever else was batched. Caught vLLM 0.26.0 returning different logprobs for byte-identical requests (vllm#51187). |
| MIRR | LLM agents diagnose a hidden fault across five noisy microservices under partial observability. Built for the Meta x PyTorch hackathon. |
| Tessera | Verified-by-default distribution for ML models and datasets. Publishers sign once; untrusted mirrors can deny a model but cannot forge one. |
| Shadowbook | A limit order book in Rust. 416ns p50 insert, 100M fuzzed ops against an oracle, zero divergences. |
| Flint | An x86-64 kernel in Rust. Ring 3 isolation proven by a breakout harness. |
More, including Thrice, Starling and CAIRN, at utkarshbahuguna.me.
I publish what the measurements say. When all-pairs comparison showed half of a Lockstep finding was an artifact, the retraction went next to the result.
Rust C++ Python CUDA/Triton Go Java TypeScript
Portfolio ·
LinkedIn ·
Email



