🧠 The AI Factory

A self-evolving multi-agent MVP generator that transforms business requirements into complete specifications.

🎯 Overview

The AI Factory uses collaborative AI agents to:

1. Parse business requirements from documents (PDF, images, text)
2. Generate user stories and architecture
3. Draft backend and frontend specifications
4. Review for quality issues
5. Fix identified problems
6. Evaluate final output with PM gatekeeper
7. Learn from failures to improve over time

✨ Key Innovation

Agents read persistent "playbooks" (Markdown files with learned rules). When output fails PM evaluation, a Coach agent extracts lessons and updates playbooks. The system gets smarter with each failure.

🚀 Quick Start

1. Install dependencies:

   poetry install

2. Set up environment variables:

   cp .env.example .env
   # Edit .env with your API keys

3. Run the application:

   poetry run streamlit run app.py

📋 Environment Variables

GOOGLE_API_KEY=<gemini_key>       # For document vision
GROQ_API_KEY=<deepseek_key>       # For agent reasoning
TAVILY_API_KEY=<tavily_key>       # For library verification
LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=<key>

🏗️ Architecture

Agents

Agent	Purpose	API
VisionAgent	Document parsing	Gemini 1.5 Pro
PMAgent	User stories, clarifications, evaluation	DeepSeek R1 (Groq)
TechLeadAgent	Architecture design	DeepSeek R1 (Groq)
DevTeamAgent	Backend/Frontend specs	DeepSeek R1 (Groq)
QAAgent	Quality analysis	DeepSeek R1 (Groq)
CoachAgent	Lesson extraction	DeepSeek R1 (Groq)

Workflow Phases

1. Ingestion - Parse uploaded document
2. Foundation - Generate user stories + architecture
3. Clarification - Resolve ambiguities
4. Drafting - Create backend/frontend specs
5. QA - Analyze for issues
6. Fixing - Apply corrections
7. Gatekeeper - PM evaluation (may loop back)
8. Output - Generate final files

Safety Mechanisms

• ⏱️ Rate Limit Protection - Exponential backoff (1s → 16s)
• 🔄 Infinite Loop Guard - Max 3 retries then force proceed
• 📝 Memory Deduplication - 80% similarity threshold
• 💾 State Recovery - Graceful corruption handling

📁 Directory Structure

ai_factory/
├── app.py                 # Streamlit entry point
├── src/
│   ├── agents/           # AI agent implementations
│   │   ├── base.py       # Abstract base + Groq wrapper
│   │   ├── vision.py     # Gemini document parser
│   │   ├── pm.py         # Product Manager
│   │   ├── tech_lead.py  # Architect
│   │   ├── dev_team.py   # Backend/Frontend
│   │   ├── qa.py         # Quality checker
│   │   └── coach.py      # Memory updater
│   ├── utils/            # Utilities
│   │   ├── state.py      # Session management
│   │   ├── files.py      # I/O operations
│   │   └── fuzzy.py      # Text similarity
│   └── schemas.py        # Pydantic models
├── memory/               # Persistent playbooks
│   ├── pm_playbook.md
│   ├── tech_lead_playbook.md
│   ├── backend_playbook.md
│   ├── frontend_playbook.md
│   └── qa_playbook.md
└── projects/             # Generated outputs

📚 Playbooks

Each agent reads from learned rules to improve over time:

• PM Playbook - Requirements quality, evaluation criteria
• Tech Lead Playbook - Architecture decisions, tech choices
• Backend Playbook - API design, error handling
• Frontend Playbook - UX patterns, state management
• QA Playbook - Security checks, quality standards

🎮 Usage Tips

1. Upload clear requirements - The better the input, the better the output
2. Review ambiguities - The system flags unclear requirements
3. Check QA report - Critical issues must be addressed
4. Trust the process - If rejected, the system learns and improves

📊 Output Files

Each project generates:

File	Contents
00_MASTER_PROMPT.md	Implementation guide
01_User_Stories.md	Detailed requirements
02_Architecture.md	System design
03_Backend_Final.md	Backend specification
04_Frontend_Final.md	Frontend specification
05_QA_Report.md	Quality analysis

🛡️ Error Handling

• Gemini fails: Falls back to text input
• Groq fails: Retries with exponential backoff
• Tavily fails: Uses fallback version list
• PM rejects: Extracts lessons and retries
• Max retries: Forces proceed with warnings

📄 License

MIT License - Feel free to modify and use.