• Designed and built multiple RL/SWE evaluation environments for frontier AI coding agents for long horizon tasks including task specs, Dockerized sandboxes, hidden/public tests, and scoring logic • Validated tasks through oracle/no-op checks and model rollouts to ensure failures reflected genuine reasoning gaps, not flaky tests or verifier ambiguity

Contributor to Karpathy/autoresearch, Google Workspace CLI, NVIDIA NemoClaw, LangChain, PyTorch, and Kubernetes: security hardening, compiler correctness, and agent reliability across codebases used by millions of developers. Karpathy/autoresearch: self-healing loop that parses its own stack traces and recovers from OOM and syntax crashes. Shopify's CEO ran it uninterrupted overnight and got 53% faster rendering from 93 automated commits (merged) Google Workspace CLI: eliminated a TOCTOU race condition leaking OAuth tokens to ~21,000 users, fixed silent auth failure where credentials never persisted while auth appeared to succeed, added Google Meet integration with deterministic requestId hashing (3 merged PRs) NVIDIA NemoClaw: enforced 0o600 on session config world-readable by default, fix propagated into 3 downstream forks within 13 hours of merge. Closed two distinct .env exposure windows on sandbox startup and remote deployment (3 merged PRs) LangChain & LangGraph: fixed silent tool-calling failures across Mistral, Groq, and Ollama causing entire agent workflows to stop on provider switch, across ~500k developers monthly (4 merged PRs) PyTorch: fixed silent data corruption in Inductor C++ kernels, any codebase mixing OpenMP with OpenCV silently produces wrong training outputs with no error thrown (approved) Kubernetes: modernizing Device Resource Allocation API across ~3M production clusters (under review)

Partnering with founders and executives to design and build custom technical solutions that solve expensive problems and create new opportunities. My work focuses on delivering clear ROI by reducing manual overhead, creating new revenue streams. REPRESENTATIVE PROJECTS: Automated client onboarding for Meteoric Boost by wiring DocuSign, Notion, and Discord via webhooks — cut manual overhead by 93%. Built Node.js scripts for Isotta Homes to parse fee structures from Google Sheets and auto-generate Xero invoices, eliminating a recurring 2-hour manual billing cycle. Deployed production booking and lead generation pages for property clients (devasramam.live, anpustay.com).

At Celer AI, my goal is to solve the #1 cause of doctor burnout: the burden of documentation. ▸ Building a tool that uses AI to convert spoken voice notes into ready-to-sign medical reports, designed to work with any existing EMR. ▸ Currently in the pre-launch phase, actively connecting with and learning from medical professionals across India to build the solution they truly need.