Building and optimizing a production voice pipeline
Real-time voice orchestration
Built a LiveKit-based STT/LLM/TTS pipeline for production voice agents, integrating tool execution, knowledge retrieval, and campaign workflows. Refined the agent tick loop and prompting to improve tool-use and retrieval reliability, alongside structured post-call extraction and sentiment analysis.
Latency optimization and model evaluation
Tuned silence detection, endpoint timing, prompt size, caching, connection reuse, and LiveKit behavior across the response path. Reduced configuration-dependent turn latency from approximately 2.2 seconds to 0.8–1.2 seconds. Evaluated model combinations against latency, response quality, and cost to select configurations appropriate to the workload.
Event-driven conversation processing
Built a separate meeting-intelligence pipeline that transformed recordings into diarized transcripts, normalized transcript chunks, and extracted topics and sentiment. Structured the processing stages around events and surfaced their outputs in dashboards, connecting conversation data to downstream product features.