VoiceOps AI
A real-time voice streaming backend that keeps latency under 200 ms with 100+ sessions talking at once.
Overview
Voice AI feels broken the moment it lags. VoiceOps AI is the backend that keeps audio moving: sessions stream in over WebSockets, Kafka consumer groups share the work, and Kubernetes adds consumers when the backlog starts to grow.
The pipeline
Each stage scales on its own:
Ingest
Clients stream audio to the backend over WebSockets.
Queue
Audio is published to Kafka, which decouples ingest from processing.
Process
Kafka consumer groups share the processing load across workers.
Scale
Kubernetes HPA adds consumer pods as load rises and consumer lag grows.
Store
PostgreSQL, MongoDB and Redis back the services.
Watch
Consumer lag and P95 latency are tracked across the streaming pipeline.
What it shows
This project is about the unglamorous half of voice AI: backpressure, horizontal scaling and observability, so the model on the other end always gets audio on time.