Sub-250 ms, or it isn't a conversation
End-to-end voice latency decides whether a call feels human. How we stream every stage to get the reply under the threshold people expect from each other.
Getting an AI agent through a payer phone tree
Payer phone menus are all different and change without warning. Why we run these calls as a state machine with tones, hold detection and a hard escalation rule.
The double-dose problem
A care agent has to ask the right question. How a symptom to drug-class map and evals on real transcripts catch the risk a fluent agent misses.
RAG that holds up in production
Hybrid retrieval, answers grounded in sources, regression evals and a model cascade. What it took to make a diligence tool a PE team could act on.
The listener constellation: one voice, a dozen models
Specialized listeners read one transcript in parallel and push context into the main model through a four-level priority queue, up to an immediate interrupt.
The nurse decides
Human supervision is the architecture of our care system, required by Medicare care management and by trust. How the rule shapes what the AI is allowed to do.
The data layer: what you own when you own the calls
Owning the operator means owning every interaction. A longitudinal record that compounds, per-patient learning, and memory built as a graph plus vectors.
The work happens while you're still talking
Background agents book, refill, notify and document during the call, six at once, without colliding. Idempotency keys keep retries from doing anything twice.
Documenting into any EHR without an integration project
Certified EHR APIs are read-only, and write access takes vendor sign-off. A browser agent that uses the EHR the way a nurse does ships in days.