🧠 Brain — Multi-Agent AI Platform

A production-shaped MERN application where a single user request is handled by five cooperating agents instead of one big prompt:

Planner → Retriever → Coder → Executor → Reviewer

Built by Ayush Kumar Gupta.

---

---

Architecture

Layer	Tech
Frontend	React 18, Redux Toolkit, React Router, Tailwind CSS, Vite
Backend	Node.js, Express
LLM	Groq API (llama-3.3-70b-versatile by default)
Agent orchestration	Custom LangGraph-style state pipeline (backend/src/agents/orchestrator.js)
RAG	Qdrant vector DB — auto-falls back to an in-memory cosine-similarity store if Qdrant isn't running
Caching / rate limiting	Redis — auto-falls back to an in-memory cache if Redis isn't running
Persistence	MongoDB (users + chat history)
Auth	JWT
Sandbox execution	vm2 (demo-grade — see note below)
Containers	Docker + docker-compose

Everything optional degrades gracefully. You only need a Groq API key and MongoDB to run this locally — Redis and Qdrant are nice-to-haves that the app detects and falls back from automatically.

How a request flows

1. User sends a message → POST /api/chat/stream (Server-Sent Events).
2. Planner breaks the request into steps.
3. Retriever pulls relevant context from the vector store (Qdrant or in-memory).
4. Coder drafts the answer, using the plan + retrieved context.
5. Executor runs any JavaScript in the draft inside a sandboxed VM.
6. Reviewer checks the draft + execution output and produces the final answer.
7. Every step streams live to the frontend, which lights up the agent pipeline visualizer node by node in real time.
8. The full pipeline result is cached in Redis for 10 minutes, keyed by a hash of the query, so repeated questions resolve instantly.

---

---

Quick start (local, no Docker)

1. Backend

cd backend
cp .env.example .env
# edit .env and set GROQ_API_KEY + JWT_SECRET (Mongo/Redis/Qdrant are optional)
npm install
npm run dev

Backend runs on http://localhost:5000.

2. Frontend

cd frontend
npm install
npm run dev

Frontend runs on http://localhost:5173 and proxies /api to the backend.

3. Use it

Open http://localhost:5173, register an account, and start chatting. Watch the pipeline visualizer light up as each agent runs.

---

Quick start (Docker — full stack incl. Mongo, Redis, Qdrant)

cp backend/.env.example .env   # then edit GROQ_API_KEY + JWT_SECRET at repo root
docker compose up --build

• Frontend: http://localhost:5173
• Backend: http://localhost:5000/api/health

---

Environment variables (backend)

Variable	Required	Default
GROQ_API_KEY	✅	—
JWT_SECRET	✅	—
PORT	❌	5000
CLIENT_ORIGIN	❌	http://localhost:5173
MONGO_URI	✅	mongodb://localhost:27017/brain
REDIS_URL	✅	in-memory fallback
QDRANT_URL	✅	in-memory fallback
GROQ_MODEL	✅	llama-3.3-70b-versatile

---

Project structure

brain/
├── backend/
│   ├── server.js
│   └── src/
│       ├── agents/         # planner, retriever, coder, executor, reviewer, orchestrator
│       ├── config/         # groq.js, db.js, redis.js
│       ├── rag/            # qdrantClient.js, vectorStore.js (fallback)
│       ├── models/         # User.js, Chat.js
│       ├── middleware/     # auth.js
│       └── routes/         # authRoutes.js, chatRoutes.js
├── frontend/
│   └── src/
│       ├── components/     # Sidebar, ChatWindow, AgentPipeline, MessageBubble, Footer
│       ├── pages/          # Login, Register, ChatPage
│       └── store/          # Redux Toolkit slices
├── docker-compose.yml
└── README.md

---

Notes on production-hardening

This repo is structured the way a production app would be, with a few things deliberately simplified for a local/demo footprint — called out here so nothing is presented as more finished than it is:

• Sandbox execution uses vm2 in-process with a 2s timeout. For real untrusted-code execution, run the Executor agent's code in an isolated container or microVM (e.g. Firecracker, gVisor), not in-process.
• Image generation mode calls Pollinations.ai, a free, open-source image API that needs no signup or key. Anonymous usage is rate-limited (~1 request/15s) with no uptime SLA — fine for demos, not for production traffic. Swap in OpenAI Images, Stability AI, or Replicate in backend/src/agents/imageProvider.js for production-grade generation.
• PPT mode renders a real .pptx via pptxgenjs in backend/src/agents/formatter.js — no external API needed.
• RAG embeddings use a deterministic hashed bag-of-words vector when no external embedding model is wired up, so the app works offline. Swap in a real embedding model for production-quality retrieval.
• Rate limiting is a simple fixed-window counter in Redis — fine for a demo, consider a sliding-window/token-bucket approach at scale.

License

MIT