The Challenge

A digital health startup was building a platform that needed to assess patient symptoms, determine urgency levels, and route cases to appropriate specialists — all within a HIPAA-compliant infrastructure. Their existing triage process was manual: nurses reviewed submissions in a queue, often taking hours to prioritise urgent cases over routine ones.

They needed an AI-powered system that could provide initial assessments in seconds, not hours — while maintaining the clinical accuracy and compliance standards that healthcare demands.

What We Did

We assembled a team of 3 engineers — a senior Python developer with NLP experience, a full-stack developer for the platform interface, and a DevOps engineer specialising in healthcare compliance. The project kicked off with a two-week discovery phase where we worked with the client’s medical advisors to define triage categories, urgency thresholds, and clinical decision trees.

The core system used retrieval-augmented generation (RAG) architecture: a vector database of medical guidelines and symptom-condition mappings, combined with LLM APIs for natural language understanding of patient descriptions. The RAG approach meant the system’s responses were grounded in verified medical knowledge, not just general language model outputs.

The Approach

HIPAA compliance shaped every architectural decision. All data was encrypted at rest and in transit. The LLM API calls were routed through a proxy that stripped personally identifiable information before sending symptom descriptions for analysis. Patient data never left the client’s AWS environment.

We built an evaluation framework that tested the system against 500+ historical cases (anonymised) reviewed by clinicians. The system needed to match or exceed nurse triage accuracy before going to production. We iterated on the prompt engineering and retrieval pipeline until accuracy consistently exceeded 92%.

The Result

The system went live in 4 months. Average triage time dropped from 2.3 hours to 3.4 minutes — a 68% reduction in time-to-assessment. Urgent cases that previously waited in a general queue are now flagged within seconds. The platform maintains 99.9% uptime and has passed two independent HIPAA compliance audits since launch. The client has since expanded the AI integration to include follow-up scheduling and care plan recommendations.