Postmortem: Building a Hindi-English Voice Agent for Fintech
anonymous_ZsP · reddit · 2026-08-07
A developer shared a detailed postmortem of building a Hindi-English code-mixed voice agent for fintech over six months.
Core Pain Points:
- Number Readbacks: In fintech, the agent often read amounts or OTPs in jarring full-English chunks or read reference numbers as single giant numbers, severely damaging user trust.
- Language-Switch Stutter: TTS engines noticeably paused or shifted accents at Hindi-English boundaries, causing users to mistake it for a scam.
- Concurrency Latency: Under real-world call concurrency, some TTS providers spiked to 800ms+ latency, ruining the call experience.
- Compliance: RBI-adjacent scrutiny and data residency rules disqualified several otherwise good options.
The Fix: Switching to a TTS provider that treated Indian code-mixing and number normalization as first-class features, and rigorously testing through the actual telephony pipe at real concurrency.
More from coding & agent
- Prime Agent: A Self-Improving RLM Agent for Coding Workflows — PrimeIntellect-ai · 2026-08-07
- Google Open-Sources Agent Skills Library for Its Product Ecosystem — google · 2026-08-07
- Semantica: Graph-Native Infrastructure for Accountable AI Systems — semantica-agi · 2026-08-07
- Clean Code Author Open-Sources swarm-forge for Coordinating AI Agents — unclebob · 2026-08-07
- Scrapling Hits 72.9k Stars: The Ultimate Scraping Backbone for AI Web Agents — techNmak · 2026-08-07
- Prime Intellect Ships Open-Source Prime Agent to Solve LLM Context Rot — blaizedsouza · 2026-08-07