FastAPI service for Bekwarra (bkv) translation, deployed on Render's free tier. Kept deliberately small: no torch. ONNX Runtime does the inference and the Marian tokenizer is pure sentencepiece, which is what makes it fit in 512 MB of RAM (measured: 672 MB with torch, ~330 MB without).

This repo exists as a Git source for Render — Hugging Face repos are git repos, so no GitHub account is needed.

Endpoints: /health, /translate, /detect