Multi-model AI chat assistant built on React 19 and TypeScript. Routes requests across Gemini 3 Flash/Pro, Kimi K2, Groq Compound, and DeepSeek V3.1 via OpenRouter, augments answers with live Exa web search and precise source citations, streams the chain of thought while it reasons. Universal file context and locally persisted history.
Live: zetaai-pink.vercel.app | Code: GitHub
Migrated a conventional chatbot to an LLM-powered conversational system for a multi-tenant hospitality mobile app (Flutter). Constrained assistant workflows with multi-guardrails, refusal behavior, and RAG retrieval; fully revamped tenant app and support flow in Figma.
Event-driven on-chain monitoring with automated alerts across EVM and non-EVM chains. Hybrid WebSocket plus polling ingestion cut API spend ~80%; Next.js dashboard cut manual monitoring 60%. Node.js, Express, Redis, MongoDB, Railway, GitHub Actions.
Rust CLI compressing Model Context Protocol responses to prevent LLM context bloat. Saves 40-75% of context tokens with zero-copy streaming. Drop-in filter for Claude Code, Cursor, MCP agent loops.
Code: GitHub
Golang reverse proxy swapping CLI/BYOK auth with Kiro/AWS credentials for Claude Code teams. Transparent HTTP/2 and WebSocket proxying under 1.5ms overhead with automatic IAM token refresh.
Code: GitHub
Sandboxed Electron AppImage unifying AI platforms on Linux. Session isolation, hotkey scratchpad, runs on Arch, Ubuntu, Fedora, Debian.
Code: GitHub