4-Week Healthcare AI MVP Cuts ER Errors 41%
The healthcare AI MVP was a 4-week build for AtriaCare's emergency department: an on-prem, fine-tuned open-source clinical LLM behind a HIPAA-aligned gateway that produces citable, auditable triage suggestions inside their VPC. Retrieval runs over their internal protocol library with citation rendering, and a clinician-in-the-loop UI records every accept or override for compliance and continuous evaluation.
On-prem deployment of a fine-tuned open-source clinical LLM behind a HIPAA-aligned gateway, retrieval over their internal protocol library with citation rendering, and a clinician-in-the-loop UI that records every accept/override for compliance and continuous evaluation.
Book a free 30-minute callHealthcare
Industry
4 Weeks
Timeline
Project Overview
1
What We Built
- On-prem deployment of a fine-tuned open-source clinical LLM behind a HIPAA-aligned gateway, retrieval over their internal protocol library with citation rendering, and a clinician-in-the-loop UI that records every accept/override for compliance and continuous evaluation.
2
The Challenge
- AtriaCare's emergency department was buckling under triage variance: junior nurses were over-admitting, seniors were over-discharging, and the LLM tools they trialled all wanted patient data shipped to a US provider. They needed an MVP that ran inside their VPC, integrated with their EHR, and produced citable, auditable triage suggestions.
3
Our Solution
- On-prem inference cluster running Llama-derived clinical model with vLLM
- Protocol-grounded RAG with sentence-level citations clinicians can verify
- Override capture pipeline feeding weekly evaluation and prompt tuning
- EHR integration via FHIR with read-only scope and audit logging
- Cost dashboard showing inference dollars per encounter