kbhuza/ithute-umunthu-zulu

isiZulu bias detection and correction pipeline.

• —Detection: Umunthu v4 n-gram classifier (sklearn / joblib).
• —Correction: tiny-aya-global GGUF base + fine-tuned LoRA adapter, served via llama.cpp.

Endpoints

• —GET / — web UI
• —GET /health — service health
• —POST /correct — { "text": "...", "lang": "zu" }
• —POST /detect-correct — runs detection then correction in one call

Notes

• —Free CPU Basic tier (2 vCPU / 16 GB RAM) is sufficient.
• —Cold-start ~30 s while the GGUF loads; steady-state ~2–3 s/request.
• —The fast-path (deterministic noun swap) returns in ~50 ms when applicable.