6 merged PRs across LLMOps and agent-framework codebases, including Agenta (YC) and Mastra (97k★+ combined across the repos below). I verify root cause before touching code, and ship fixes maintainers don't have to rewrite.
🔨 Right now: 10 PRs in review. LiteLLM router and proxy work: cooling down deployments on stream-fetch failure (#42815), an opt-in flag to honor Retry-After across multiple deployments (#34408), capping upstream error bodies at 64 KB (#34392). Plus a Redis streams worker fix in Agenta (#7587).
| Repo | Scale | What I actually fixed |
|---|---|---|
| LiteLLM (LLM Gateway) | 62.1k★ | In review (8 PRs): router cooldown and reporting on fetch_stream() failure (#42815); opt-in flag to honor Retry-After with multiple deployments, with docs (#34408, litellm-docs#643); cap on upstream error-response bodies at 64 KB (#34392); tool-schema unpacking budget for Bedrock/Vertex (#34391); end-user spend counter reseeding (#34298); hanging-request alert alias fix (#34307); MCP semantic tool filter format fix (#32606); postcss bump for GHSA-r28c-9q8g-f849 (#34533). |
| Mastra (Agent Framework) | 29.1k★ | Merged: removed leaked sub-agent reporting instructions from BUILD/Explore mode prompts (#24862); fixed dictation transcripts being dropped on normal stop in the React client (#20016). |
| Agenta (LLMOps, YC) | 4.9k★ | Merged: JSON Diff evaluator silently ignoring case_insensitive_keys for any key with uppercase characters (#7161); trace output column now shows the exception message when outputs is null (#4747); correct breadcrumb label for SDK evals (#4619). In review: Redis streams worker that hung forever on a silent connection, timeout and health check added (#7587). Also built a RAG chatbot example (LangChain LCEL + vector retrieval + Agenta tracing), live at agenta-ai-examples.vercel.app. |
| TraceRoot (Observability) | 1k★ | Merged: improved how errors are displayed when testing a bring-your-own-key connection (#1426). |
Synthetic Oracle — ML stock signal engine XGBoost + RF + 4 more models, 74% accuracy, 542 NSE stocks, 13-layer decision engine, SHAP explainability, Groq LLM analysis. Live · Code
CodeArena — Docker-sandboxed coding platform Judge0 execution, real-time submissions, AI code review via Groq/Llama. Live · Code
StartupOS — Ops automation dashboard GitHub API + Google Sheets + Slack webhooks, rule-based alert engine, multi-tenant JWT auth. Live · Code
Languages: Python · C++ · TypeScript · JavaScript · SQL
Backend: FastAPI · Node.js · Express · asyncio · REST APIs · JWT
Data & infra: MongoDB · PostgreSQL · Redis (Streams) · Docker
AI / ML: LangChain · LiteLLM · XGBoost · scikit-learn · SHAP · Groq (Llama)
Testing & tooling: pytest · Vitest · Ruff · Git · GitHub · uv · pnpm
Frontend & deployment: React · Vercel · Netlify
📩 Open to remote backend / AI infra roles.
