Jetson Smart Board Recorder - edge capture with local RAG

Whiteboard notes captured, read and made searchable entirely on-device - no cloud vision API, no per-frame billing, nothing leaving the room.

Jetson Nano 4GBYOLOv8n + TensorRTLlama 3.1 8B + ChromaDBSolo build

The system

A CSI camera feeds GStreamer; a YOLOv8n TensorRT model watches for the gesture or person cue that triggers capture; homography rectifies the board; GPU EasyOCR reads it; SQLite keeps it; ChromaDB and a local Llama 3.1 8B answer questions over it.

My role

Sole contributor - architecture, the multi-threaded vision pipeline, TensorRT export and OCR optimisation, the local RAG and SQLite backend, Docker packaging, and the Flask Mission Control dashboard.

Result

Moving OCR off the CPU took a frame from 4-6s to about 1.2s, while the trigger model holds 30+ FPS in real time on a 4GB board. Search, Markdown notes, PDF export and semantic chat all run offline.

4-6s to 1.2s per OCR frame30+ FPS trigger inferenceZero cloud cost, fully on-prem

Capture pipeline - all on device

CSI cameraGStreamer ingestYOLOv8n TensorRT triggerHomography rectifyGPU EasyOCRSQLite

Retrieval layer

OCR textChromaDB embeddingsLocal Llama 3.1 8BFlask Mission Control

Repository ↗