AI Software Engineer

A multi-agent system that takes a plain-English description of a web app and writes the full HTML/CSS/JS code for it. Built with LangGraph and Groq-hosted LLaMA 4 models.

Live demo: https://ai-software-engineer.onrender.com

How it works

Three agents in a LangGraph pipeline:

User prompt → Planner → Architect → Coder → Generated app

1. Planner — decides the app structure (single-file vs multi-file, complexity tier)
2. Architect — breaks the plan into specific implementation tasks (max 3 tasks for simple apps, 1 for complex ones)
3. Coder — writes the actual code for each task, file by file, using only HTML/CSS/JS (no frameworks, no external APIs)

The generated files are written to generated_project/ and served live via FastAPI — you can view the output in the browser immediately.

Stack

• langgraph — agent orchestration (StateGraph with Planner → Architect → Coder nodes)
• langchain-groq — LLM calls via Groq API (LLaMA 4 Scout / Maverick / LLaMA 3.3 70B with automatic fallback on rate limits)
• fastapi + uvicorn — serves the chat UI and generated output
• pydantic — structured output schemas for agent state

Run locally

pip install -r requirements.txt

Add your Groq API key to a .env file:

GROQ_API_KEY=your_key_here

uvicorn app:app --reload

Open http://localhost:8000, type what you want to build, and watch it generate.

Project structure

agent/
  graph.py      # LangGraph StateGraph definition
  prompts.py    # System prompts for each agent
  states.py     # Pydantic state schemas
  tools.py      # File I/O tools (write_file, read_file, list_files)
app.py          # FastAPI server + chat endpoint
templates/      # Frontend UI for the chat interface
generated_project/  # Output — the app generated by the agents

Notes

• Output is constrained to vanilla HTML/CSS/JS — no npm, no build steps, runs fully offline
• Model pool tries LLaMA 4 Scout → Maverick → LLaMA 3.3 70B in order on rate limit errors