Case study · WellSky · 2021 — 2024
Document intelligence for healthcare workflows
NLP pipelines that turn referrals, forms and clinical documents into structured, searchable data.
Senior AI / Full-Stack Engineer — pipeline architecture
- Healthcare documents processed a year
- 1.5M+Healthcare documents processed a year
- Reduction in manual data extraction
- 68%Reduction in manual data extraction
- Processing throughput after redesign
- 3.2×Processing throughput after redesign
Context
A large share of healthcare information still arrives as documents. Staff spent hours reading them and re-typing the important fields into the system of record — slow, repetitive work with real consequences when something was missed.
Approach
- Designed a pipeline that classifies each document, extracts the fields that matter for that document type, and attaches a confidence score to every value.
- Routed low-confidence extractions to human review rather than writing them straight into records, so automation reduced work without hiding errors.
- Rebuilt the processing layer around asynchronous workers, which increased throughput by 3.2×.
- Exposed results through FastAPI services, improving average API response times by 40%.
Results
- The pipeline processed more than 1.5 million healthcare documents a year.
- Manual data extraction effort fell by 68%, and classification workflows reduced manual review requirements by 42%.
What I took from it
In regulated domains the review path is part of the product. Designing for “the model isn’t sure” from day one is what made the automation trustworthy enough to roll out.
Stack
- Python
- PyTorch
- Scikit-learn
- NLP
- FastAPI
- PostgreSQL
- Async workers