Sub-250ms, or it isn't a conversation
Why end-to-end voice latency is the whole game — and how to get under the threshold where a call starts to feel human instead of like a walkie-talkie.
Read ↗Teaching an AI to beat a payer phone tree
IVRs are hostile, all different, and change without warning. Why we treat them as a control problem with a state machine — not a chat problem.
Read ↗The double-dose problem
Making an agent ask the right question, not just answer one — and the evals against real transcripts that prove it caught the thing that mattered.
Read ↗RAG that doesn't fall over in production
Grounding every claim to a source, an eval harness that catches regressions, and pushing the hot path onto cheaper models without losing quality.
Read ↗The listener constellation: one voice, a dozen AIs
A swarm of specialized listeners watching one transcript in parallel, each able to inject context through a four-level priority queue — up to an immediate interrupt.
Read ↗The nurse is the law
Human-supervised AI as architecture, not a compliance checkbox bolted on at the end — and why the constraint is what makes the rest possible.
Read ↗The data layer: what you own when you own the calls
Own the operator, own every interaction — a longitudinal record that compounds, plus per-patient fine-tuning on what actually works.
Read ↗The work happens while you're still talking
Background agents that book, refill, notify, and document — silently, mid-conversation, six at once, without colliding.
Read ↗Documenting into any EHR without an integration project
A browser agent that uses the EHR the way a nurse does — no API, no vendor sign-off. Days, not quarters.
Read ↗