🤖 M.A.D.E. — Multi-Agent Data Engine

M.A.D.E. is an autonomous multi-agent pipeline built with LangGraph, FastAPI, and FAISS. It uses a graph-based state machine to orchestrate specialized LLM agents for code generation, security auditing, and sandboxed script execution backed by strict Human-In-The-Loop (HITL) guardrails.

---

🌟 Features

• Multi-Agent Collaboration:
  • 🔍 Researcher Agent: Formulates Python libraries and technical methodology for given tasks[cite: 3].
  • 💻 Coder Agent: Translates research methodology into clean, executable Python code[cite: 3].
  • 🛡️ Reviewer Agent: Performs security checks, inspects syntax, and yields audit reports[cite: 3].
• Human-In-The-Loop (HITL) Gate: Execution pauses automatically after code review, waiting for explicit user confirmation before touching the local workspace.
• Sandboxed Execution: Runs approved code inside a local, isolated workspace (generated_workspace/) with a 15-second subprocess timeout to prevent infinite loops[cite: 4].
• Vector Store Integration: Built-in FAISS vector database initialization using HuggingFace embeddings (all-MiniLM-L6-v2) for local documentation context.

---

---

🛠️ Project Structure

.
├── build_db.py         # Initializes and saves the FAISS vector database[cite: 2]
├── graph.py            # LangGraph state graph & agent node definitions[cite: 3]
├── main.py             # FastAPI REST endpoints & subprocess execution engine[cite: 4]
├── state.py            # TypedDict defining the shared agent state tape[cite: 6]
├── requirements.txt    # Python dependencies[cite: 5]
├── .env                # Environment variables (OpenRouter API Keys)
├── faiss_index/        # Local FAISS vector storage folder[cite: 1, 2]
├── web/                 # React + Vite source for the control center UI
└── frontend/            # Built UI output (served by main.py) — generated, do not hand-edit

---

🖥️ Control Center UI

The dashboard is a React app (source in web/) built with Vite and served by FastAPI as static files from frontend/. After changing anything under web/src/, rebuild:

cd web
npm install   # first time only
npm run build # writes frontend/index.html + frontend/assets/

main.py serves frontend/index.html at / and mounts frontend/ at /static — no backend changes are needed after a rebuild. For iterative UI work, npm run dev inside web/ starts a Vite dev server on port 5173 that proxies API/websocket calls to a FastAPI instance running on 127.0.0.1:8000 (see web/vite.config.js).

---

🔐 Security & configuration

M.A.D.E. is intentionally hardened so a technical judge can't poke a hole in the pipeline's logic. What you'd want to set in a real environment:

env var	default	what it controls
OPENROUTER_API_KEY	(required)	the OpenRouter key used by every LLM node
MADE_API_KEY	(unset → dev mode)	client-facing shared secret; requests to /execute-task, /approve-and-run, /reject-task must send it as X-API-Key. If unset, the server boots in DEV MODE with a loud warning and no auth.
MADE_ALLOWED_ORIGINS	http://localhost:5173,http://127.0.0.1:5173,http://127.0.0.1:8000,http://localhost:8000	comma-separated CORS allowlist. * is refused.
MADE_RATE_LIMIT_PER_MIN	12	token-bucket cap per (route, client IP)
MADE_MAX_HEAL_ATTEMPTS	3	self-heal loop retry cap; beyond this the pipeline terminates with failure_class="exhausted" instead of spinning forever
MADE_SANDBOX_TIMEOUT_SEC	10	wall-clock budget for the sandbox subprocess
MADE_MODEL_NAME	openrouter/free	the model the Coder/Researcher/Reviewer use
MADE_FAISS_INDEX_DIR	faiss_index	where build_db.py writes the vector store and graph.researcher_node reads it
MADE_RESEARCH_TOP_K	4	how many methodology docs the researcher retrieves per task

Security posture (what the modal claims — and now actually enforces)

• AST audit, in security.py, is the real security boundary. Every piece of generated code goes through it before either the Reviewer LLM sees it or the HITL gate can approve it. It fails-closed on SyntaxError, refuses imports outside a small allowlist (math, statistics, numpy, pandas, sympy, scipy, sklearn, matplotlib, and stdlib safe modules), refuses eval/exec/compile /__import__, refuses os.system/subprocess.*/socket/requests regardless of how the module was renamed, and refuses the __class__.__base__.__subclasses__ sandbox-escape family. An LLM commentary layer runs afterwards but cannot override a BLOCK.
• The same audit runs on the /approve-and-run gate, so a tampered client cannot POST forged code past the Reviewer.
• /ws/logs broadcasts to one queue per connection so concurrent viewers all see every line.

Real RAG

build_db.py builds a FAISS index over a curated methodology corpus (pandas, numpy, sympy, scipy, sklearn recipes) using all-MiniLM-L6-v2 embeddings. researcher_node retrieves the top-K relevant recipes for each task and grounds the methodology prompt in them. The retrieved snippets are exposed to the UI as "grounded in:" citations. Missing index → graceful fallback to an unguided LLM prompt with a warning.

First-time setup

pip install -r requirements.txt
python build_db.py               # builds faiss_index/ (downloads the embedding model once)
cd web && npm install && npm run build && cd ..
export OPENROUTER_API_KEY=…       # your key
export MADE_API_KEY=…             # any long random string; clients send it as X-API-Key
uvicorn main:app --host 127.0.0.1 --port 8000

Verify the hardening: python tests/smoketest_hardening.py (server must be running; the test sets X-API-Key=smoketest-key-987654321, so export MADE_API_KEY=smoketest-key-987654321 in the terminal that runs uvicorn).