TuyizereMandella/virixyai

Voice ai-agent infrastructure that turns a Kinyarwanda and other African Languages phone call into a completed public-service transaction. # Virixy AI Voice-agent infrastructure that turns a Kinyarwanda phone call into a completed public-service transaction. **The claim this project proves:** the hard, undifferentiated part of African-language voice AI (speech recognition, speech synthesis) is already solved and open-sourced (Digital Umuganda for Kinyarwanda). What's missing is the layer that turns that into an agent that *does something* — holds conversation state across turns, decides what to ask next, and calls a real backend action. That orchestration layer is what this repo builds, and it's designed so a new language means adding a new STT/TTS backend, not rewriting the engine. ## Demo scenario: report a lost national ID Caller dials in, speaks Kinyarwanda. The agent collects four fields one at a time (full name, district, ID number if known, callback phone), confirms them back, then writes a record into the mock civil-registry DB and reads back a reference number. A confirmation SMS follows. A live dashboard shows the record appear in real time — that's the visible proof the call "did something," not just "said something." ## Architecture ``` caller --dial--> Africa's Talking Voice --webhook POST--> this server (/voice/callback) | Say + Record (XML response) | caller speaks, AT posts recordingUrl back to /voice/callback | download audio -> STT (Modal: Whisper) | orchestrator: slot-filling, intent, session state | writes to registry DB (SQLite) visible on /dashboard | TTS (Modal: Chatterbox) -> Play + Record next turn (XML response) | on completion: SMS confirmation via Africa's Talking SMS API ``` Voice and SMS both go through Africa's Talking now (moved voice back from Twilio - Twilio only sells international, not local, Rwandan numbers, meaning every test call would cost an international rate off MTN/Airtel; Africa's Talking has real in-country Rwanda voice presence). Africa's Talking's ` ` is record-then-respond, not full-duplex streaming — the conversation is turn-based by design (call …