A clinical-grade triage agent across four regions.

The brief.

A US healthcare operator running 47 clinics across four regions needed to triage symptom-intake forms in seconds. Speed mattered — patients in the queue wanted to know whether to wait, drive to an ER, or book a routine appointment. But the failure mode that could not happen was missing a high-acuity case. A delayed cardiac symptom misrouted to “routine” was the kind of outcome that ended companies.

The operator had tried two off-the-shelf vendors. Both produced systems that worked in the demo and quietly degraded once real intake forms — with their typos, code-switching and missing fields — started flowing through.

Devmint partnered with the operator's clinical leadership and engineering team to ship the agent in production — with the eval, the audit log, and the on-call rotation that an HIPAA-bound clinical system actually requires.

What we shipped.

• Routing · 018-route specialty triageCardiology, urgent care, primary, dermatology, ENT, orthopedics, pediatrics, OB/GYN. Each intake form is classified against the routing graph with confidence scores, not a yes/no flag.
• Safety · 02Hard-coded escalationSpecific symptom patterns — chest pain, stroke indicators, pediatric fever above thresholds — escalate to a nurse-on-call within 15 seconds. No LLM judgement involved on these; the rules are deterministic.
• Compliance · 03On-prem PHI inferencePHI-bearing fields are processed on the operator's own infrastructure with a locally hosted model. No identifiable data leaves the operator's network. The cloud LLM only sees de-identified structured features.
• Audit · 04Replay-ready audit logEvery decision, every retrieval, every model version is logged. A compliance officer can reconstruct any single triage decision against a newer model — the audit log was a precondition for the contract.

How we built it.

The framing we landed on with the operator was that the agent is a triage assistant, not a triage decision-maker. Every escalation, every routing recommendation is reviewed by a nurse-on-call before it reaches the patient. The agent's job is to do the speed work — the parsing, the classification, the initial routing — and the human's job is to confirm.

That framing is what made the engagement defensible. It also informed the eval. We didn't evaluate the agent against “correct decision” — we evaluated it against “nurse-on-call sees this in <30 seconds with the right context attached.” The 62% intake speed improvement is measured against that target, not the headline.

The eval harness ran against 8,000 historical cases, each reviewed by a board-certified clinician. Every PR ran the full eval. New regions could not go live until they hit 99.4% on the high-acuity correctness floor.

The stack.

Outcomes.

Across six months of production, the agent has triaged 47,302 intake forms with zero PHI incidents, an 11.2% escalation rate to nurse-on-call (the design target was 8–12%), and a 62% reduction in time-to-routing against the pre-launch baseline.

The audit log was used in a formal HIPAA review three months after launch and passed without findings. Devmint continues on a quarterly retainer running the operate-and-improve programme.