All case studies
Active DevelopmentSpeech · Realtime2025 — present
Voxera
Low-latency duplex voice layer for conversational agents — designed against walkie-talkie feel.
Role · Lead engineer
Overview
A duplex voice layer targeting sub-300ms perceived latency in agentic conversations.
Problem
Stitching together STT, LLM, and TTS naively yields conversations that feel like walkie-talkies — long pauses, no interruption, awkward turn-taking.
Solution
Streaming pipelines with early endpoint detection, interruption handling, and partial-response synthesis.
Tech stack
- TypeScript
- WebRTC
- Streaming STT/TTS
Engineering notesWhat's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Engineering notes
What's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Design goals
- Perceived latency measured, not assumed.
- First-class interruption — user can talk over the agent.
- Backpressure handled cleanly across the STT → LLM → TTS chain.
Philosophy
Voice UX lives or dies on latency and interruption. Everything else is a footnote.
Planned architecture
WebRTC transport; streaming STT with endpoint detection; token-streamed LLM; partial TTS synthesis with an interruption channel.
Current stage
Transport layer and latency measurement scaffolding.
Links
Repository and demo links will appear here once the project is ready for public review.