back to work
AI · Multi-agent
Quanta Terminal
Personal Project
A production-grade multi-agent AI terminal that deploys five specialist agents in parallel to generate structured investment memos for quantum computing and AI chip stocks, streamed live to the browser.
LangGraphFastAPIFastMCPNext.js 14TypeScriptPythonGemini 2.5SupabasePostgreSQLRedisServer-Sent EventsRailwayVercelLangSmithSentry
Responsibilities
- Architected and shipped a full-stack AI research platform end-to-end — agent pipeline, MCP servers, REST API, SSE streaming, auth, caching, and deployment.
- Designed a LangGraph fan-out pipeline that runs five specialist AI agents concurrently, each with its own MCP-connected data source, then synthesises all outputs into a single investment memo via Gemini 2.5.
- Built five FastMCP stdio servers exposing tools for SEC EDGAR, Yahoo Finance, NewsAPI, and arXiv; enabling agents to call live, structured data sources through the MCP protocol.
- Implemented real-time Server-Sent Events streaming so users see agent progress and memo content appear token-by-token as the pipeline executes.
- Deployed a production system on Railway (FastAPI backend) and Vercel (Next.js frontend) with Supabase for passwordless auth and PostgreSQL persistence, and Upstash Redis for caching and rate limiting.
Key Actions & Decisions
- Chose LangGraph's StateGraph over a sequential chain to enable true parallel agent execution — all five agents run via asyncio.gather, reducing total pipeline latency compared to running them sequentially.
- Adopted the MCP protocol for all data access rather than inline API calls — each data source is an isolated subprocess with a defined tool schema, making sources independently swappable without touching agent logic.
- Used Redis atomic INCR+EXPIRE for rate limiting with a fail-open design so infrastructure issues never block legitimate users.
- Verified JWTs from Supabase using ES256 public-key verification with JWKS caching, with HS256 as a fallback; ensuring auth works correctly across both magic-link and session-refresh flows.
Result
Deployed and live. Generates a structured, multi-section investment memo covering SEC filings, earnings trends, analyst consensus, news sentiment, and academic research for any of 10 tracked tickers in under 2 minutes. Supports authenticated research chat with each specialist agent post-memo, with full conversation memory per thread.