# Project: Web Development Swarm (MAS)
**Status:** Phase 1 (Data Engineering) Complete | Phase 2 (Model Training) Pending

## 🎯 Objective
Build a local Multi-Agent System (MAS) for web development.
- **Orchestrator:** 26B Model (running on GPU) to plan, route, and review.
- **Specialists:** 7-8B Models (running on CPU/System RAM) fine-tuned for specific domains (Frontend, Backend, DevOps, etc.).

## 🛠 Technical Architecture
- **Hardware Target:** RTX 4090 (GPU) for Orchestrator; System RAM (CPU) for Specialists.
- **Fine-Tuning Method:** QLoRA via **Unsloth** (for speed and VRAM efficiency).
- **Base Model:** Llama-3-8B-Instruct.
- **Communication:** Structured JSON/Manifest-based "Blackboard" architecture to prevent context drift.

## ✅ Completed Work
- [x] Project directory structure initialized: `~/projects/web-dev-swarm/`
- [x] **Frontend/UI Specialist Dataset** created:
    - **File:** `~/projects/web-dev-swarm/datasets/frontend/raw/frontend_gold_standard.jsonl`
    - **Content:** 50 high-quality instruction/response pairs (React + Tailwind CSS).
    - **Categories:** `component_gen`, `layout`, `refactor`, `a11y`, `state_ui`.

## 🚀 Immediate Next Steps (Phase 2: Training)
1. **Environment Setup:** Create a dedicated Python environment with `unsloth`, `torch`, and `xformers`.
2. **Training Execution:** 
    - Write a training script to apply LoRA adapters to Llama-3-8B using the `frontend_gold_standard.jsonl` dataset.
    - Execute training and save the LoRA adapters.
3. **Verification:** Run a "blind" test on the fine-tuned model to ensure it adheres to the specialized UI/UX patterns.

## 📂 Key Paths
- **Project Root:** `~/projects/web-dev-swarm/`
- **Dataset:** `~/projects/web-dev-swarm/datasets/frontend/raw/frontend_gold_standard.jsonl`