# Dev-AI: Fine-Tuned 27B Code Review Model

**Goal:** Train a 27B model on software development data, run it on CPU as a code review partner.

## Architecture

```
┌──────────────┐    ┌──────────────┐    ┌──────────────┐
│  Dataset     │───▶│  QLoRA Train │───▶│  Merge +     │
│  (Code, SO,  │    │  (CPU, 4-bit)│    │  llama.cpp   │
│  Docs, PRs)  │    │  ~14 days    │    │  Server      │
└──────────────┘    └──────────────┘    └──────────────┘
                                              │
                                              ▼
                                        ┌──────────────┐
                                        │ Code Review  │
                                        │ API (port    │
                                        │ 8081)        │
                                        └──────────────┘
```

## Workflow

1. Vincent writes code → 2. POST to fine-tuned model → 3. Model reviews/corrects → 4. Vincent iterates

## Hardware

- **Training:** Ryzen 9 9900X CPU, 60GB RAM (no GPU)
- **Inference:** llama.cpp server, ~17GB RAM for Q4 27B model
- **GPU (4090):** Reserved for Vincent's own inference, not touched

## Status

- [ ] Dataset collection
- [ ] Training script with checkpoint resumption
- [ ] systemd background service
- [ ] Model merge + GGUF conversion
- [ ] llama.cpp server setup
- [ ] Code review integration workflow