Healthcare · Helix Health

AI triage assistant deployed across 38 clinics

Cut intake time by 41% with HIPAA-compliant LLM workflows.

AI triage assistant deployed across 38 clinics
Services
  • AI / ML
  • Custom Software
  • QA
Stack
  • Python
  • Anthropic
  • FastAPI
  • Pinecone
  • AWS
Published

January 28, 2026

41%
Intake time reduction
38
Clinics live
>95%
Eval accuracy on prod traffic
0
PHI incidents
Challenge

The starting point

Helix Health's intake process was paper-heavy and inconsistent across clinics. Wait times were the #1 patient complaint.

Earlier LLM pilots had stalled on HIPAA review — no evals, no eval harness, no story for hallucination control.

Approach

How we worked

We built an eval-first triage assistant: every intent classification was scored against a clinician-labeled gold set before any feature shipped.

PHI never left the VPC — retrieval and inference run on AWS with private endpoints and explicit logging.

The first launch was shadow-only: the assistant ran in parallel with human intake for six weeks, with weekly clinical reviews.

Results

What changed

Intake time dropped from an average of 14 minutes to 8 minutes — without a single PHI incident.

Patient NPS lifted 18 points across the 38 pilot clinics.

Helix's clinical team now owns the eval harness, with new evals added weekly.

Have a similar project?

We've shipped this kind of work before — let's talk about how the lessons apply to yours.

Posted 2026-01-28 by Official Byte. Read more on case studies.