Healthcare Intelligence Platform
Re-architecting a healthcare data platform supporting 1M+ patient records — FHIR/EHR ingestion, RAF scoring, and AI-driven patient insights.
1M+
Patient Records
FHIR / EHR
Focus
RAF
Scoring
Architecture
Overview
As a senior engineer at a healthcare AI startup, I am leading the re-architecture of a data platform that supports 1M+ patient records. The goal is an AI-driven healthcare intelligence layer — natural-language patient insights on top of reliable, normalized clinical data.
Problem
The existing platform could not support AI analytics at the required scale. Ingestion pipelines and data models needed to be redesigned so multi-provider EHR data could be normalized, scored, and queried without becoming a bottleneck for downstream intelligence features.
Architecture
EHR/FHIR sources feed redesigned ingestion pipelines, then a canonical normalization layer, RAF score computation, and optimized PostgreSQL/ETL workflows that power AI analytics. See the interactive diagram above.
Challenges
- Scale: 1M+ patient records require ingestion and query paths that stay performant as volume grows
- Multi-provider FHIR/EHR: different source systems must land in a consistent patient model
- RAF computation: risk scores need to stay current as clinical data arrives
- AI analytics: ETL and PostgreSQL have to support natural-language insight workloads, not just operational reporting
Scale / Performance
- Platform redesigned around 1M+ patient records
- PostgreSQL and ETL workflows optimized for high-performance AI analytics
- Ingestion pipelines rebuilt so analytics is not blocked by source-system variance
Interesting Engineering Decisions
FHIR as the interchange layer
Normalizing multi-provider EHR data through FHIR keeps the canonical model independent of any one vendor's schema, which is required for both RAF scoring and AI insights.
Redesign ingestion before adding AI features
Natural-language insights fail if the underlying records are inconsistent. The re-architecture starts at ingestion and data models, then layers RAF and AI on a stable foundation.
Lessons Learned
- AI products in healthcare are data-platform problems first
- RAF scoring is only as good as the normalized clinical record behind it
- Query and ETL optimization have to be designed for analytics load, not just ingestion throughput
Outcome
In progress (Mar 2026 – Present): leading the platform re-architecture so 1M+ patient records can power FHIR-normalized ingestion, RAF scoring, and AI-driven patient insights.