Technology

Gemini Embedding 2

Multimodal embedding model that maps face images into 3072-dimensional vectors. Each reference face is pre-processed and stored — giving us the representational power of a frontier model without the corporate guardrails.

Face Detection

Gemini 2.5 Flash vision model detects faces and returns bounding box coordinates. Sharp crops the face region with padding, isolating it from background noise before embedding.

Top-K Nearest Neighbors

Cosine similarity computed against all 1,691 reference embeddings. Top 30 matches are averaged — a genuine match sustains high similarity across many faces, not just one or two by chance.

Sigmoid Scoring

Raw similarity mapped to 1–100 via a steep sigmoid curve (midpoint 0.722, steepness 200). Small differences in embedding similarity produce clean separation between matches and non-matches.

1,691 Reference Embeddings

Pre-computed dataset loaded into memory at startup (~64MB). Scoring is pure CPU math — cosine similarity across all references in under 50ms. The bottleneck is the Gemini API, not us.

Guardrail-Free Architecture

Embeddings give you the power of frontier models in a form you own. Extract vectors, build your own scoring — no content policy between you and your application. The big models are incredible. Their wrappers are not.