- •Built a production voice interview agent using Pipecat, AWS Bedrock Nova Sonic, and LiveKit/WebRTC, migrating from LiveKit Agents SDK + GPT-4o Realtime to improve latency and reliability.
- •Designed concurrent interview slot management using Redis sorted sets and Lua scripts for atomic operations, eliminating race conditions under load.
- •Scaled backend to multi-EC2 deployment with BullMQ for distributed job dispatch and worker orchestration.
- •Implemented observability pipeline with Grafana, Loki, Promtail, and Langfuse (via OpenTelemetry) for real-time monitoring and LLM tracing.
- •Resolved production issues including Whisper hallucination during silence, S3 multipart upload credential errors, and WebRTC VAD concurrency bugs.
NestJSPipecatAWS BedrockLiveKit/WebRTCRedisBullMQGrafana/LokiLangfuseS3