OBSERVA
OBSERVA is an offline-first, on-device Android assistant for blind and low-vision users that provides real-time hazard-aware guidance, scene understanding and accessibility feedback using private edge AI without sending camera data to the cloud.
Core trick
Runs a lightweight on-device vision model continuously on camera frames to give blind users hazard-aware audio guidance with no cloud calls.
Description
$27
README excerpt
# OBSERVA **OBSERVA is an offline-first AI vision assistant for blind and low-vision users.** Existing AI vision apps assume the user can already aim a camera at what matters. For a blind user, that *is* the problem. OBSERVA reverses the assumption: a lightweight, always-on, on-device model gives **continuous ambient awareness** of the surroundings first, then helps the user orient toward what's there. ## Promises - **Offline-first** — the safety-critical core runs entirely on-device and works in airplane mode. Cloud is an optional enhancement, never a requirement. - **Privacy-first** — camera imagery is processed locally and does not leave the device for any core function. - **Accessibility-first & safety-first** — designed for non-visual use, built to never be confidently wrong about hazards. Android-native, low-latency, battery-conscious, on-device inference (ExecuTorch). ## Current status (v3.1.0 — NPU-accelerated, voice control everywhere) **Working today, verified on a Galaxy S25 Ultra (Snapdragon 8 Elite) in Airplane Mode:** - **Real on-device ML object detection on the Hexagon NPU** — YOLOv8n (COCO-80) exported to **ExecuTorch** (320×320), running on the **Qualcomm Hexagon NPU (HTP v79)** at **~2–3 ms** (`forward backends=[QnnBackend]`, `QNN/NPU ACTIVE`) — ~10–15× faster than the XNNPACK CPU path, which stays as the automatic fallback. Bundled as `assets/models/observa_detector.pte` and run locally; person/vehicle/large-obstacle detections drive the hazard engi