README.md

---
title: Jeeva Embedding Service
emoji: 🐕
colorFrom: red
colorTo: yellow
sdk: docker
pinned: false
---

# Jeeva Embedding Service

Single-purpose FastAPI service that accepts a dog photo and returns two 512-dim embedding vectors:
- `face_embedding` — derived from the top 40% crop of the image (heuristic face region)
- `body_embedding` — derived from the full image

Uses [MegaDescriptor-T-224](https://huggingface.co/BVRA/MegaDescriptor-T-224) from the `wildlife-tools` package.

## Endpoints

**POST /embed**
- Accepts: `multipart/form-data` with a `file` field (image)
- Returns: `{ face_embedding: float[], body_embedding: float[], dims: 512 }`

**GET /health**
- Returns: `{ status: "ok", model_loaded: bool, last_embed_ts: float | null }`

## Notes

- First request after a cold start will be slow (~60–90s on CPU) while the model loads
- All inference runs on CPU — no GPU required
- Designed for POC scale; not optimised for throughput