Case study · 2024 to present
A voice-agent platform for a Fortune-5 healthcare buyer
A platform-based blueprint for cost-effective, production-grade voice agents: telephony, orchestration, LLM and TTS swaps, evaluation, and the unglamorous reliability work that makes them stick in regulated environments.
Senior Forward Deployed Engineer
Stack
Python · Twilio / SIP · FastAPI · Kubernetes · LLM orchestration · Streaming TTS/ASR · Evals
Outcomes
- ~95% containment on outbound flows in production deployments.
- Telephony, agent runtime, and evaluation layered into a reusable platform so new agents ship in days, not quarters.
- Public write-up of the architecture pattern on the CVS Health Tech Blog.
What I owned
Technical leadership of the voice-agent platform: the runtime that hosts agents, the telephony and streaming layer that connects them to real callers, the orchestration patterns the agents use to do work, and the evaluation framework that keeps quality from drifting under us. I worked alongside platform, data, and clinical-ops stakeholders to take this from prototype to a thing real customers were routed through.
What shipped
A re-usable platform (not a one-off bot) that decouples the conversation policy from the LLM, the LLM from the TTS, and the TTS from the telephony provider. Each layer is swappable. Outbound deployments reach ~95% containment without escalation; new agents ship in days against the same primitives instead of starting from a blank Twilio app. The public version of this pattern is the CVS Health Tech Blog post linked above.
Lessons
Voice is a latency budget problem first, an ML problem second, and a policy problem always. The win was making boring things, turn-taking, barge-in, retries, fallback grammars, into platform primitives so feature teams could spend their time on the conversation, not the plumbing.