I work in the gap between AI research and production: reinforcement-learning environments, agent infrastructure, and the systems plumbing that makes them trustworthy.
- 🔭 Built RL environments and LLM infrastructure at Scaler AI Labs, delivered to frontier-lab clients (OpenAI, xAI, Meta)
- 🔧 3 merged PRs in HuggingFace TRL, including a silent training-corruption bug in GRPO that I found, reported, and fixed
- ⚙️ For fun I write systems software: a storage engine in Go, developer tools, and application security
- 🎓 CS undergrad at BITS Pilani, studying at Scaler School of Technology
- 🌐 More about me at gowthamsai.in
| Contribution | What it was |
|---|---|
trl#6534 merged |
GRPO silently accepted a reward function returning the wrong number of rewards, broadcasting one value across every completion and poisoning training. Filed issue #6533 with a repro, shipped the guard plus a regression test, replicated it in RLOO on review. |
trl#6503 merged |
Defaulted GRPO to the importance-sampling-corrected KL gradient. A maintainer pushed back, so I reproduced their script, found my own "no-op" claim was wrong, and said so. |
trl#6477 merged |
Completed the GRPOConfig.loss_type docs across all 8 loss types. |
lm-evaluation-harness#3950 open |
The 67-subject TMMLU+ benchmark stopped loading on datasets>=4. Repointed it at a parquet-native mirror, proving parity by diffing git blob hashes across 267 files. |
| Project | What it is |
|---|---|
| entangle | Hand a live coding-agent session to another machine, or another person, with every file path rewritten so it resumes. Three-word code, end-to-end encrypted, no server in the middle. Claude Code, Codex, opencode. Go |
| Viveka | An RL environment that teaches agents to pause and confirm before an irreversible action, graded by a 6-component deterministic reward with no LLM-as-judge. Top 9 of 70,000+ at the Meta × Scaler OpenEnv hackathon. Python · GRPO/TRL |
| gravel | An embeddable LSM-tree key-value storage engine written from scratch: skip-list memtable, WAL with crash recovery, SSTables, bloom filters, leveled compaction. ~295K writes/sec. Go |
| oncall-env | "SWE-bench for incident response": 48 scenarios over 11 real root-cause types, under partial observability so the agent has to investigate. Python |
- Scaler AI Labs, AI Engineer (Jan–May 2026): delivered 5 LLM agent environments to frontier-lab clients (Asana MCP at a perfect 1.0 across 52 tasks, HelloSign at 92.94% client QA), a distributed MCP server at sub-200ms p99, and Pulse, an LLM extraction pipeline live for a 20,000-user org.
- Wealthy, SDE-1, promoted from intern (Aug 2024–Aug 2025): owned the internal CRM used by 7,700+ partners behind ₹4,000+ Cr AUM; cut API latency 32% and lifted conversion 37%.
I like building things that don't exist yet. If it sounds hard, I'm probably already working on it.