AutoGenesis is a multi-agent autonomous AI developer platform that transforms natural language prompts into complete, production-ready software projects — including source code, unit tests, CI/CD pipelines, and Dockerfiles — all with zero manual intervention.
It uses a quality-gated, multi-pass pipeline with four specialized AI agents that plan, code, review, and deploy. The system gets smarter over time via an XP-based intelligence system and RLHF feedback loop.
Describe your idea, pick a template, or speak it — AutoGenesis does the rest.
View all generated files with syntax highlighting and an AI-assigned quality score.
See your generated web project rendered live in a sandboxed preview pane.
Track the AI's growth stages — from Baby to Sage — as it learns from every project.
| Feature | Description |
|---|---|
| 🏗️ Multi-Agent Pipeline | Architect → Coder → Reviewer → DevOps agents working in sequence |
| 🔁 Quality-Gated Loops | Automatically retries and fixes until quality thresholds are met |
| 📡 Real-Time Streaming | Live progress updates via Server-Sent Events (/run-stream) |
| 🧠 XP & Skill Tree | Gamified AI progression: Baby → Child → Teen → Adult → Expert → Sage |
| 💬 Voice Input | Speak your project idea via Web Speech API |
| 🎯 Live Preview | Instantly preview generated web projects in an iframe sandbox |
| 🔬 RLHF Feedback | Star ratings captured and converted to Oumi-compatible DPO training data |
| ⚙️ DevOps Generation | Auto-generates unit tests, GitHub Actions CI, and Dockerfiles |
| 🚀 One-Click Deploy | Deploy to Vercel or build & push a Docker image straight from the UI |
| 🗂️ Persistent Memory | Remembers past projects and uses context for smarter future generations |
| 🔐 Security Hardened | CORS allowlist, trusted host middleware, security headers, authenticated /reset |
| 🌊 Kestra Integration | AI memory summarization workflows triggered on milestones |
┌────────────────────────────────────────────────────────────┐
│ AutoGenesis Pipeline │
│ │
│ User Prompt │
│ │ │
│ ▼ │
│ 🏗️ Architect Agent ──→ Project Plan (files, tech stack) │
│ │ │
│ ▼ │
│ 💻 Coder Agent ──→ Source files (multi-language) │
│ │ │
│ ▼ │
│ 🔍 Reviewer Agent ──→ Quality score + issue list │
│ │ │
│ ▼ │
│ ⚙️ DevOps Agent ──→ Tests, CI/CD, Dockerfile │
│ │ │
│ ▼ │
│ 🎯 Quality Gate ──→ Pass / Retry loop │
│ │
└────────────────────────────────────────────────────────────┘
| Priority | Provider | Model | Role |
|---|---|---|---|
| 1 | CCAPI | MiniMax-M2.5 | Primary gateway |
| 2 | Gemini | gemini-2.5-flash | Planner & Reviewer |
| 3 | Groq | llama-3.3-70b | Coder & DevOps |
All providers include exponential backoff retry, rate-limit detection, telemetry, and automatic failover.
- Python 3.9+ · FastAPI · Uvicorn
- Gemini API (Architect, Reviewer) · Groq API (Coder, DevOps)
- Server-Sent Events for real-time streaming
- Oumi RL for RLHF fine-tuning pipeline
- Next.js 16 · TypeScript · Tailwind CSS
- Web Speech API for voice input
- Docker · Docker Compose · GitHub Actions
- Vercel deployment · Kestra AI workflow automation
- CodeRabbit for automated PR reviews
- Python
3.9+ - Node.js
20.x - npm
>=9 - Optional:
vercelCLI,docker
# Copy and fill environment variables
cp .env.example .env
cp backend/.env.example backend/.envMinimum required keys:
| Variable | Description |
|---|---|
GEMINI_API_KEY |
Primary planner/reviewer |
GROQ_API_KEY |
Primary coder/devops |
NEXT_PUBLIC_API_URL |
Frontend → backend URL (e.g. http://localhost:8000) |
Top-Tier Mode (optional):
TOP_TIER_MODE=true
TOP_TIER_MAX_ATTEMPTS=3
TOP_TIER_QUALITY_THRESHOLD=0.85
TOP_TIER_MIN_REVIEW_SCORE=0.80
TOP_TIER_MIN_UI_SCORE=0.70
AGENT_PROVIDER_MAP={"architect":"gemini","coder":"groq","reviewer":"gemini","devops":"groq"}
ALLOW_MOCK_FALLBACK=falsecd backend
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS/Linux
# source .venv/bin/activate
pip install -r requirements.txt
uvicorn api:app --reload --host 0.0.0.0 --port 8000cd frontend
npm install
npm run devOpen http://localhost:3000 in your browser.
docker-compose up --build| Endpoint | Method | Description |
|---|---|---|
/run |
POST |
Non-streaming project generation |
/run-stream |
POST |
Streaming generation via SSE |
/preview/start |
POST |
Start live preview server |
/preview/status |
GET |
Preview server status + logs |
/preview/stop |
POST |
Stop live preview server |
/optimize |
POST |
Optimize a prompt before generation |
/fix |
POST |
AI-assisted code fix |
/explain |
POST |
AI code explanation |
/memory |
GET |
Project memory history |
/intelligence |
GET |
XP / level status |
/skills |
GET |
Skill tree data |
/feedback |
POST |
Submit RLHF star rating |
/feedback/stats |
GET |
RLHF & training dataset stats |
/deploy |
POST |
Deploy to Vercel or Docker |
/status |
GET |
Provider status + pipeline telemetry |
/kestra/summarize |
POST |
Trigger Kestra memory workflow |
/kestra/status |
GET |
Kestra integration history |
/reset |
DELETE |
Admin reset (requires x-admin-token) |
Set VERCEL_TOKEN in your environment, then:
# From UI: click Deploy → Vercel
# Or via API:
POST /deploy { "target": "vercel" }Set
VERCEL_ALLOW_INTERACTIVE=trueto use interactive CLI login locally.
POST /deploy { "target": "docker", "push_image": true }AutoGenesis captures user star ratings and converts them to training data:
| File | Purpose |
|---|---|
storage/feedback.json |
Raw star ratings |
storage/oumi_dataset.jsonl |
Oumi-compatible DPO dataset |
storage/rlhf_meta.json |
Training metadata |
Auto-training controls:
RLHF_AUTO_TRAIN=trueRLHF_MIN_FEEDBACK_FOR_TRAIN=10
AutoGenesis can trigger Kestra workflows for AI memory summarization:
KESTRA_ENABLED=true
KESTRA_API_URL=http://localhost:8080
KESTRA_NAMESPACE=autogenesis
KESTRA_FLOW_ID=summarize_memory
KESTRA_AUTO_SUMMARIZE=true
KESTRA_SUMMARY_INTERVAL=10 # trigger every N memory savescd backend
pytest -qcd frontend
npm run test:smokecd frontend
npm run buildautogenesis/
├── assets/ # Screenshots & media
├── backend/
│ ├── agent/
│ │ ├── agent.py # Provider routing + retry
│ │ ├── orchestrator.py # Multi-pass quality pipeline
│ │ ├── agents.py # Agent class definitions
│ │ ├── coder.py # Language-aware code generation
│ │ ├── reviewer.py # Code review + scoring
│ │ ├── quality_gate.py # Multi-signal quality evaluation
│ │ ├── sandbox.py # Subprocess execution sandbox
│ │ ├── preview.py # Runtime preview management
│ │ ├── memory.py # Project memory persistence
│ │ ├── feedback.py # RLHF feedback loop
│ │ └── intelligence.py # XP / progression tracking
│ ├── api.py # FastAPI endpoints + SSE
│ └── oumi_rl.py # Oumi RL integration
├── frontend/
│ └── src/app/page.tsx # Main Next.js UI
├── cli/
│ └── autogenesis.py # Cline CLI
├── kestra/
│ └── flows/summarize_memory.yaml
├── .coderabbit.yaml # CodeRabbit PR review config
├── docker-compose.yml
├── vercel.json
└── .env.example
See CONTRIBUTING.md for guidelines.
This project is licensed under the MIT License.
Built for Hackathon 2025 · Made with ❤️ and a lot of AI