XaeroAI

AI runtime for Cyan whiteboard-to-diagram pipeline.

Pipeline

Whiteboard Photo → YOLO (shapes) → PaddleOCR (text) → Dictionary (correct) → Phi-3 (mermaid)

Models

Model	Purpose	Size	Format
cyan-sketch	Shape detection	25MB	ONNX
paddleocr	Text recognition	7.5MB	ONNX
cyan-lens	Mermaid generation	2GB	GGUF Q4

Models: scripts/models/

Usage

cargo run --release --bin test_pipeline -- \
  --models-dir scripts/models \
  --image whiteboard.jpg \
  --verbose

Output

Detected Shapes (14):
  [0] rectangle  conf=0.87  "START"
  [1] diamond    conf=0.70  "Valid?"
  [2] rectangle  conf=0.84  "Process"
  ...

Generated Mermaid:
flowchart TD
  A[START] --> B{Valid?}
  B -->|Yes| C[Process]

Project Structure

xaeroai/
├── src/
│   ├── lib.rs           # Library exports
│   ├── pipeline.rs      # Main pipeline orchestration
│   ├── runtime.rs       # ONNX/GGUF model loading
│   ├── arrow_detector.rs # Line/arrow detection
│   ├── dictionary.rs    # OCR correction
│   └── skill.rs         # Model metadata
├── scripts/models/
│   ├── whiteboard-detector/
│   ├── paddleocr/
│   └── cyan-lens/
└── config/
    └── ocr_dictionary.json

Configuration

OCR dictionary in config/ocr_dictionary.json:

{
  "min_length": 3,
  "max_distance_short": 1,
  "max_distance_long": 2,
  "terms": {
    "flow_control": ["START", "END", "YES", "NO"],
    "components": ["USER", "SERVER", "API", "DATABASE"]
  }
}

Dependencies

• ort - ONNX Runtime
• llama-cpp-2 - GGUF inference
• image / imageproc - Image processing

License

Business Source License 1.1