Bridging Digital Identity to the Physical World

Inspiration

We've solved identity online—profiles, bios, social graphs, shared interests. But the moment we step into the physical world, context disappears. Access becomes awkward. Discovery feels invasive.

The problem isn't information. It's how and when it appears.

We built a system that brings digital context into physical spaces—recognizing who someone is and surfacing relevant information at the right moment.

The Pipeline

Our system chains multiple APIs to go from camera feed to identified person:

Overshoot AI — Detection
- RealtimeVision SDK analyzes camera feed
- Detects when someone is intentionally engaging (not just passing by)
- Checks: centered in frame, facing camera, stationary
- Requires 3 consecutive confirmations to reduce false positives
MediaPipe + OpenCV — Feature Extraction
- Extracts facial landmarks from captured image
- Converts face into a numerical feature vector
Wood Wide AI — Embedding & Matching
- POST /api/models/embedding/train — Trains embedding model on face database, returns model_id
- GET /api/models/{model_id} — Poll until status is COMPLETE
- POST /api/models/embedding/{model_id}/infer — Generates embedding vectors
- Compares embeddings using Euclidean distance to find the closest match
Output — ChatGPT API
- Returns identified person's name for next steps (display greeting, log access, etc.)