🔥 ButaAI

Autonomous On-Device Engineering Squad & Decentralized Swarm LoRA Engine

100% Local • Apple Silicon Metal GPU-Accelerated • Zero Cloud Dependency • Federated Weight Merging

---

⚡ What is ButaAI?

ButaAI is an on-device autonomous software engineering studio that operates 100% locally on your machine across Windows 10/11, macOS (Apple Silicon), and Linux. Instead of sending your private code, database schemas, and intellectual property to cloud AI APIs (OpenAI, Anthropic, Google), ButaAI turns your workstation into a full software development agency powered by a 6-specialist autonomous engineering squad.

Furthermore, through Decentralized Swarm LoRA & P2P 16MB Chunk Distribution, millions of ButaAI developers collectively improve one unified open-weights model by training lightweight parameter delta adapters from validated unit tests without sharing any proprietary code.

---

🖥️ Cross-Platform Hardware Acceleration Matrix

ButaAI auto-detects your platform and dynamically switches acceleration backends:

OS / Platform	Hardware Acceleration	Training Engine	Fuser & Quantizer Backend
macOS (M1–M4)	Apple Silicon Metal GPU (MPS)	Apple MLX / Metal MPS	MLX Fast-Fuse + 4-Bit Unified Memory
Windows 10/11	NVIDIA GeForce RTX (CUDA 12.x+)	PyTorch + CUDA Wheels	PyTorch PEFT + BitsAndBytes 4-bit / GGUF Q4_K_M
Windows 10/11	AMD Radeon / Intel Arc	DirectML (torch-directml)	FP16 DirectML Merge
Windows / Linux	Multi-Core x86_64 CPU	AVX-512 / OpenMP PyTorch	CPU GGUF Quantization

---

🥊 ButaAI vs. Cloud Frontier Models

Why is an on-device specialized squad superior to monolithic cloud LLMs for software development?

Feature	Cloud Monoliths (Claude 3.5 Sonnet / GPT-4o)	ButaAI Autonomous Studio
Data Privacy	❌ Code and schemas sent to remote cloud data centers	✅ 100% On-Device. Local GPU RAM execution, zero telemetry
Running Cost	❌ $15 – $30 per 1M tokens ($100s/mo in API bills)	✅ $0.00 Forever. Free local GPU computation
Cognitive Architecture	❌ Single generalist prompt approves its own code	✅ 6 Isolated Specialists with strict adversarial QA veto
Offline Operation	❌ Zero functionality without internet	✅ 100% Offline. Code on planes, trains, and air-gapped setups
Model Evolution	❌ Static proprietary weights updated twice a year	✅ Decentralized Swarm LoRA. Collective weights merge continuously
P2P Distribution	❌ Giant multi-gigabyte cloud bottlenecks	✅ 16 MB P2P Chunks. Merkle-verified BitTorrent / peer LAN sync
Deliverables & Live Canvas	❌ Disjointed markdown blocks	✅ Interactive Live Canvas. Mobile/Tablet frames, code copy & .ZIP export
Multimodal Assets	❌ Cloud token fees for images & video	✅ Native On-Device Creation. Vector SVGs, Flutter animations, 18s kinetic launch video storyboards

---

🎨 Multimodal Creation Engine (Zero Cloud Token Fees)

ButaAI includes native on-device tools to craft complete creative and engineering assets:

• 🖼️ Procedural Vector SVGs (generate_image): Generates crisp, infinite-resolution app icons, badges, UI mocks, and architecture diagrams directly on disk with specular lighting, concentric beveled borders, and instant web previews.
• ✨ 60 FPS Motion & Physics (generate_animation): Synthesizes production-ready Flutter CustomPainter widgets with AnimationController and hardware-accelerated CSS keyframe animations.
• 🎬 18-Second Launch Video Storyboards (generate_video_storyboard): Produces high-retention launch video blueprinted into 5 cinematic acts (Hook, Problem, Reveal, Proof, CTA) with JSON timeline data, interactive HTML5 web player, and automated FFmpeg render scripts.
• 📱 Live Multi-Agent Interactive Canvas: Live interactive preview workspace directly inside the Cockpit featuring mobile (340x640), tablet (460x600), and fluid desktop frames with responsive reactive state testing and code inspection.

---

👥 The 6 Autonomous Specialists

When you pitch a feature or complete app concept, ButaAI orchestrates 6 dedicated engineering specialists who debate requirements and self-verify before writing production code:

1. 🎩 Buta CEO: Executive scoping, milestone planning, and release roadmaps.
2. 📋 Product Architect: Rigorous PRDs, user stories, edge cases, and acceptance criteria.
3. 📐 System Architect: Clean Architecture, SQLite entity relational schemas, and state management patterns.
4. 🎨 Lead Frontend Engineer: Modern design systems, liquid glass styling tokens, responsive layouts, and zero-leak controllers.
5. ⚙️ Lead Backend Engineer: High-performance SQLite repositories, batch transactions, and data pipeline caching.
6. 🛡️ QA & Security Verifier: Automated unit test suites, mock generation, memory leak audits, and vulnerability guardrails.

---

🐝 Decentralized Swarm LoRA & P2P Chunk Distribution

Traditional AI training requires hundreds of millions of dollars in centralized compute. ButaAI introduces Decentralized Swarm Weight Merging via GitHub Actions & P2P Swarm:

[Developer A: Windows RTX 4090] ── Verified 100% Tests ──> [Adapter ΔWA] ──┐
[Developer B: MacBook M3 Max]   ── Verified 100% Tests ──> [Adapter ΔWB] ──┼─> [P2P 16MB Chunk Swarm] ──> [Master Model]
[Developer C: Linux Titan GPU]  ── Verified 100% Tests ──> [Adapter ΔWC] ──┘

1. Local Gradient Generation: When you build an application and the QA Verifier confirms a 100% green test pass, your local engine calculates a lightweight rank-16 LoRA adapter in under 60 seconds on Apple Silicon Metal or Windows NVIDIA CUDA.
2. 16 MB BitTorrent / P2P Swarm Chunks: Large 35 GB–40 GB adapters are sliced into 16 MB content-addressed blocks with cryptographic SHA-256 Merkle root verification, enabling zero-cloud LAN/peer sync.
3. 4-Bit Model Fuser: Locally merges adapters into the base Ollama model and exports a quantized GGUF (buta-ai:colossus-fused) for lightning-fast memory-efficient inference.
4. Zero Privacy Leakage: Only mathematical delta weights ($\Delta W$) and test-pass assertions are synchronized; your proprietary code never leaves your machine.

---

🚀 Quickstart

🪟 Windows 10/11 Installation

Option A: 1-Line PowerShell Bootstrap (Run as Administrator or User):

iwr -useb https://raw.githubusercontent.com/bekovrafik/ButaAI/main/install.ps1 | iex

Option B: Double-Click Batch Installer:

1. Clone the repo: git clone https://github.com/bekovrafik/ButaAI.git
2. Double-click install.bat to automatically install dependencies and configure NVIDIA CUDA / DirectML.
3. Double-click launch.bat anytime to launch the Gateway and open the Cockpit!

---

🍏 macOS (Apple Silicon M1–M4) & Linux

Option A: 1-Line Automated Install:

curl -fsSL https://raw.githubusercontent.com/bekovrafik/ButaAI/main/install.sh | bash

Option B: Manual Setup:

# 1. Clone repository
git clone https://github.com/bekovrafik/ButaAI.git
cd ButaAI

# 2. Install dependencies & build Gateway
npm install
npm run build

# 3. Launch with 1-click script
./Start\ ButaAI.command

Open your browser at http://localhost:8080 to access the Cockpit Studio!

---

📦 Project Structure

ButaAI/
├── app/                        # Flutter Web & Desktop Cockpit Studio
│   ├── lib/screens/            # Interactive Cockpit UI & Streaming Chat
│   ├── lib/widgets/            # Live Canvas, Specialist Avatars, Deliverables & Instincts
│   └── lib/services/           # Local Gateway, Hardware & Metal/CUDA Connectors
├── src/                        # High-Performance Node.js / TypeScript Gateway
│   ├── agents/specialists.ts   # 6 Autonomous Specialist Prompts & Personas
│   ├── core/engine.ts          # Multi-Turn Tool Calling & Execution Loop
│   ├── tools/media_tools.ts    # Multimodal Generator (Vector SVGs, Flutter Animations, 18s Video)
│   ├── swarm_hub.ts            # Decentralized LoRA Consensus & Adapter Sync
│   └── swarm_p2p.ts            # P2P 16MB Chunk Distribution & Merkle Engine
├── training/                   # Cross-Platform Local ML Training & Swarm LoRA
│   ├── data/                   # Verified JSONL Instruction & Multimodal Datasets
│   ├── generate_dataset.py     # High-Purity Synthetic Dataset Synthesizer
│   ├── train.py                # Cross-Platform QLoRA Trainer (Metal MPS & NVIDIA CUDA)
│   ├── fuse_quantize.py        # 4-Bit Model Fuser (MLX & PyTorch BitsAndBytes)
│   ├── swarm_merger.py         # FedAvg Parameter-Space LoRA Merger & Safety Auditor
│   └── github_trainer.py       # GitHub Swarm Synchronizer & Delta Contributor
├── .github/workflows/          # CI and GitHub Swarm Training Release Pipeline
│   ├── ci.yml                  # Engine & Flutter Build Verification
│   └── buta_swarm_train.yml    # Decentralized Swarm LoRA Train & Release Bot
├── docs/                       # Architecture & Training Manifestos
│   └── AI_ARCHITECTURE_AND_TRAINING.md
├── install.sh                  # macOS / Linux 1-Click Installer
├── install.ps1                 # Windows 10/11 PowerShell Bootstrap Installer
├── install.bat                 # Windows Double-Click Batch Installer
├── launch.bat                  # Windows 1-Click Service Launcher
└── Start ButaAI.command        # macOS 1-Click Double-Click Launcher

---

🧪 Local Training & GitHub Swarm CLI

1. Synchronize Collective Master Weights from GitHub

python3 training/github_trainer.py --sync

Downloads the verified consensus master LoRA adapter from GitHub Releases (swarm-weights-latest), audits cryptographic SHA-256 integrity, and activates it on your Apple Silicon Metal GPU.

2. Train Local LoRA (Apple Silicon Metal & Windows NVIDIA CUDA)

# macOS (Apple Silicon Metal MPS)
python3 training/train.py 20

# Windows 10/11 (NVIDIA CUDA or CPU)
python training/train.py 20

3. Fuse & Quantize Local Weights into Ollama (4-Bit GGUF)

# Merges the 35GB-40GB collective adapter into Ollama with 4-bit quantization
python3 training/fuse_quantize.py --tier colossus --quantize 4bit

# Run the fused model immediately
ollama run buta-ai:colossus-fused

4. Contribute Verified Delta to the Collective Swarm

python3 training/github_trainer.py --contribute

Audits your trained weights for privacy (guaranteeing zero personal code or prompt text leakage), packages a verified contribution_manifest.json, and guides you to open a PR for collective FedAvg synthesis on GitHub Actions!

5. Synthesize Fresh Synthetic Multimodal Datasets

python3 training/generate_dataset.py

---

🤝 Contributing

We welcome contributions from engineers, researchers, and creators worldwide!

• Check out CONTRIBUTING.md for guidelines.
• Submit issues using our Bug Report or Feature Request templates.
• Share your custom LoRA weights via the LoRA Adapter Proposal template.

---

📄 License

ButaAI is licensed under the Apache License 2.0.
Created and maintained with ❤️ by Rafik Bekov and the Butalabs team.