← Back to Projects
🤖

Production Multi-Agent Chatbot

GitHub ↗
The Business Problem

Users need to switch between separate tools for document search, database queries, and live data lookups — there's no single place to ask a question and get a trustworthy answer.

The Solution

A unified chat interface backed by four specialist agents (RAG, NL2SQL, Realtime, Memory) orchestrated by a LangGraph Supervisor, with a security gateway and evaluation layer in front of every response.

The Result
  • Rated 9.5/10 on production-readiness — deployed to AWS ECS Fargate with rolling updates and circuit breakers
  • Four specialist agents run in parallel via Map-Reduce instead of one generalist agent doing everything serially
  • Every response scored by RAGAS with a faithfulness target above 0.85, traced end-to-end in LangSmith
  • Security gateway enforces JWT auth, PII detection, and per-user rate limits before any query reaches an agent
Architecture

User Query -> Security Gateway (JWT/PII/rate-limit) -> Supervisor Agent -> Specialist Agents (parallel) -> Output Guardrails -> RAGAS Evaluation -> Response

⚙️ Technical Details For engineers & technical reviewers — click to expand

Rated 9.5/10 on production-readiness — deployed to AWS ECS Fargate with rolling updates, circuit breakers, and adversarial test suites, not left as a notebook demo. Unifies document search, database querying, and live API calls behind a single chat interface, so users stop switching between tools. Four specialist agents (RAG, NL2SQL, Realtime, Memory) are orchestrated by a LangGraph Supervisor using ReAct loops and parallel Map-Reduce execution. A FastAPI security gateway enforces JWT auth, PII detection, and per-user rate limits before any query reaches an agent, and every response is scored by RAGAS (faithfulness >0.85 target) and traced end-to-end in LangSmith.

LangGraph OpenAI GPT-4o FastAPI Pinecone DuckDB PostgreSQL Redis Celery Streamlit LangSmith RAGAS Docker AWS ECS Fargate GitHub Actions

Have questions about this project? Ask my AI assistant for details.

Ask AI about this →