19 / Real Estate & Construction
Voice AI Infrastructure & Automated QA
We fixed a property developer's sales-call voice bot so it interacts naturally, then built an auditor that scores every single call against the client's own quality rules instead of a lucky sample.
- Client
- Undisclosed (Under NDA)
- Stack & approach
- Real EstateVoice AIRealtime StreamingLLM-as-JudgeMulti-Tenant RLS
The problem
Pre-sales ran on a hosted voice agent carrying hundreds of calls a day, and the media path worked against it: a transcode broke barge-in, pre-warming stacked audio buffers into spikes, and token refresh caused auth failures. QA was manual: 2 auditors grading a fraction of volume, so an agent quoting a price that didn't exist was caught only by chance.
Solution & architecture
Fix the media path, then judge every call. A gateway carries native 16 kHz audio end-to-end with the transcode deleted and a local voice-activity detector cutting interruptions at the gateway. An audit engine scores every call against the client's indicator framework, with a reasoning trace behind each number, reading the same versioned inventory the agent reads, so a spoken price is checked against what was true at call time.
Why it is built this way
Per-tenant isolation
Force row-level security on rubric and inventory
Every call judged
Not a 2% manual sample
Grounded in truth
Price checked against call-time inventory
