Banking voice agent
A production speech agent: listen, understand, answer, in under two seconds.
Gen AI Architect · team lead
Voice is the hardest interface to get right in banking: recognition errors compound into wrong answers, every second of latency is felt, and every turn has a cost. The agent had to be accurate, fast, safe and affordable at the same time.
How it works.
STT → LLM → TTS pipeline
Speech-to-text, a language model and text-to-speech run as one production pipeline on Sber Platform V.
Fine-tuned GigaChat 2 MAX
The language layer is a fine-tuned GigaChat 2 MAX model, adapted to banking conversations and intents.
Better recognition
Speech-recognition word error rate came down from 46% to 22%, halving the errors that reach the model.
Guardrails and a CI quality gate
Guardrails keep responses inside policy, and a CI quality gate blocks any release that regresses quality.
The agent runs in production at 92% intent accuracy, 1.8 s p95 latency and ₹0.48 per turn, with recognition errors cut by more than half.
My role
- Led the team that built the voice agent
- Shaped the STT → LLM → TTS architecture, guardrails and CI quality gate
Stack
- STT
- GigaChat 2 MAX (fine-tuned)
- TTS
- Sber Platform V
- Guardrails
- CI/CD