Overview
Tahaffuz is an AI assistant for vaccination field workers in Pakistan. A vaccinator speaks a question in Urdu and gets a natural spoken Urdu answer, grounded in the program's actual SOPs and guidelines — not generic model knowledge.
What it does
- Urdu in, Urdu out — speech recognition tuned for Urdu queries, with replies pinned to the user's language and delivered as natural-sounding Urdu TTS.
- RAG over program content — answers are grounded in the full set of program modules, so guidance matches official policy.
- Built for the field — an Expo Android app designed for low-friction use on mid-range phones; the admin stays a clean, English-only Filament panel for program staff.
Stack
Laravel + Filament backend, Expo (React Native) Android app, Gemini for reasoning, embeddings, and TTS. Deployed on a VPS.
My role
End-to-end: backend, mobile app, the speech pipeline, RAG design, and deployment. The interesting problems were language ones — keeping the model answering in clean natural Urdu, and making retrieval work over Urdu/English mixed content.