FinTech · Conversational AI
AI voice agent for FinTech support
Automating high-volume tier-1 support with low-latency conversational AI and secure ledger integration.
- Tier-1 calls automated
- 60%
- Wait time
- 15m → 0s
- Ops cost reduction
- 45%
- Round-trip latency
- <800ms
Problem
A scaling European FinTech was drowning in tier-1 support calls. Their IVR was rigid, users hated it, and peak trading hours meant 15+ minute waits and high abandonment.
They needed a human-like voice agent that could authenticate users, retrieve balances securely, and complete basic transactions — without opening the core ledger to a black-box vendor.
Approach
Event-driven voice pipeline on Twilio, with Google Cloud Speech-to-Text for streaming transcription and Deepgram TTS for natural responses.
Orchestration on AWS (API Gateway + Lambda) with a fine-tuned LLaMA-based intent model and Redis for conversational state.
Account actions went through hardened internal APIs to the bank ledger — never through the model’s context window as a free-for-all toolbelt.
Latency budget owned end-to-end: STT → LLM → tool calls → TTS kept under ~800ms so the conversation felt continuous, not robotic.
Outcome
Within the first month, 60% of tier-1 calls were handled without a human. Wait times collapsed to zero for automated intents. Call-centre cost dropped ~45%, and human agents focused on disputes and edge cases instead of balance lookups.
Stack
- Twilio Voice
- Python / FastAPI
- AWS Lambda
- Google STT
- Deepgram TTS
- Fine-tuned LLaMA
- Redis