I build AI agents that ship and get used.
Top 15 of 70,000+ at the Meta × Hugging Face × PyTorch hackathon · Creator of llmswap (41K+ installs)
14 years of production engineering, lately focused on safe, auditable autonomous systems.
- Top 15 of 70,000+ at the Meta × Hugging Face × PyTorch OpenEnv Hackathon (solo, work shortlisted by Meta's AI team)
- Top 50 Finalist at the Anthropic + Accel Hackathon (invited to Anthropic India Dev Day)
- llmswap: 41K+ PyPI downloads, 5 external contributors, a Homebrew tap, and its own docs site
- 2× Show HN on Hacker News for open-source work
Krelvan · self-hostable · event-sourced · open registry
Describe an outcome in plain English and get a fully owned, running multi-agent system. A signed, append-only event ledger is the runtime, so what you see is exactly what executed; agents crash and resume with no action repeated. Prompt-injected instructions can only shrink an agent's powers, never expand them. Deny-by-default execution with typed capability grants, approval gates and secret isolation, plus an open Git-based registry of installable agents, capabilities and connectors.
krelvan.com · GitHub · Registry · Product Hunt
llmswap · 41K+ installs
Provider-agnostic LLM SDK + CLI for 11 providers (OpenAI, Anthropic, Gemini, Groq, Watsonx, Ollama, xAI, Sarvam…) with zero vendor lock-in. Workspace memory, universal tool-calling, an MCP client, and cost visibility.
Inkling · camera-first · no account needed
A child photographs a drawing and gets back a playable browser game built from their own lines, colors, and wonderfully wobbly style. Two lanes underneath: deterministic Phaser templates parameterized by a generated GameSpec, so there is always a playable game with no model in the loop, plus model-written behavior patches admitted only after passing a sandbox validator. Four-day sprint, built for OpenAI Build Week.
Selected projects
- Sampark OS · autonomous 4-agent broker negotiating fish prices for Kerala fishermen (Claude Vision + WhatsApp/Telegram)
- Eklavya Council · expert AI personas debate a question in structured rounds; a synthesis engine produces decisions, dissent, and action items. Not a chatbot, a thinking environment.
- OpenStack MCP Server · natural-language infrastructure management over MCP
- More at sreenathmenon.com/labs
I write plain-language, deeply-illustrated explainers on how modern AI actually works, aimed to be understandable from college level to senior engineer.
- The Harness: Why the Model Is the Smallest Part of Your AI Agent
- How AI Agents Actually Work
- MCP: The Port That Let AI Finally Touch the World
- From Vibe Coding to Spec Kit: How to Tell an AI What to Build
- Embeddings: How AI Knows Two Things Mean the Same
- Context Engineering: Why AI Forgets, and How We Fight It
- Ollama: Running Real AI on Your Own Machine
More at sreenathmenon.com/blog.
AI/ML: LLMs · RAG · AI Agents · MCP · Embeddings · Reinforcement Learning · LangChain · CrewAI · Vector DBs · Ollama
Working at the seam between natural language and systems, safe, auditable, reversible.