Projects & experiments

The operating system for autonomous work.

I built Krelvan around a simple belief: AI agents should become reusable software that people can install, connect, extend and genuinely own, not temporary prompts trapped inside closed platforms. A user describes an outcome, and Krelvan turns it into a working multi-step agent system with the right capabilities, tools, memory, schedules and approval controls. Its open registry provides an expanding ecosystem of installable agents, capabilities and connectors, while support for multiple model providers lets every user choose how their intelligence runs. Behind the experience is a signed, replayable execution ledger that makes every decision and action inspectable. Krelvan is more than an agent builder. It is an open foundation for creating, distributing and operating trusted AI systems on infrastructure controlled by the people who depend on them.

• TypeScript
• Next.js
• Event-sourced
• MCP

krelvan.comGitHub

Provider-agnostic LLM SDK + CLI supporting 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.

• Python
• TypeScript
• MCP

PyPIllmswap.orgGitHub

An RL environment for AI-assisted clinical triage for India's 1.07M frontline ASHA health workers. Multi-agent, adaptive curriculum, GRPO-trained to 0.947 peak reward. Built for the Meta × Hugging Face × PyTorch hackathon: Top 15 of 70,000+.

• Python
• OpenEnv
• GRPO · Qwen3

GitHubHuggingFace

An autonomous AI broker for India's informal maritime economy: a 4-agent swarm that negotiates fish prices for Kerala fishermen. Claude Vision grades catch from photos; agents broadcast over WhatsApp/Telegram, reject predatory bids, and close deals.

• TypeScript
• Claude Vision
• Multi-agent

GitHub

An AI thinking environment, not a chatbot. Expert AI personas debate a question in structured rounds; a synthesis engine produces decisions, dissent, open questions, and action items.

• TypeScript
• Anthropic Claude

npmDemo

An AI-powered Web Application Firewall for LLM apps: multi-layered defense against prompt injection, jailbreaks, and adversarial inputs. Scored 90/100 across two judges at the Deriv AI hackathon.

• TypeScript
• Next.js

Demo