Voice AI in production. On-premise, low-latency, and measured.
PhD in privacy-preserving speech (Inria / University of Lille). Co-creator of the VoicePrivacy Challenge, the community benchmark for voice anonymization. Co-founder of Nijta, where we deploy speech systems inside regulated environments, mostly public safety, transport police and healthcare.
Most of my production work lives behind customer NDAs. This account is where I rebuild it in the open, on public models and public data.
A four-part series on what it actually takes to run a voice agent in production:
- Latency lab. A cascaded voice agent (LiveKit Agents, Deepgram, Cartesia) with barge-in, semantic turn detection and SIP telephony, instrumented end to end. Where do the 800ms go?
- On Kubernetes. The same agent deployed to GKE with Terraform, plus the on-premise variant with zero egress. Real-time audio breaks most of the assumptions a normal web service makes.
- Evaluation harness. Simulated callers, scored on task completion, tool correctness, groundedness, turn-taking failures, latency and cost. Run in CI.
- Agents and compliance. Multi-agent with tool calls and warm human handoff, built in both LangGraph and Google ADK, with real-time PII redaction in the pipeline.
- 20+ peer-reviewed publications, 1,500+ citations, h-index 15 (Jul 2026) · Google Scholar
- Reviewer for NeurIPS, ICLR and Interspeech
- Organizing Committee, ISCA SPSC Symposium
- Previously: Microsoft Research (code-switched Hindi-English ASR), Inria