The problem
Scanned and handwritten medical records, superbills, and claims created substantial manual work. Information had to move from inconsistent documents into structured systems before billing workflows could proceed.
The challenge was an operational pipeline: capture the information, structure it, and make it usable inside existing healthcare workflows.
What I owned
I also built a separate end-to-end patient–provider EHR assistant, spanning 3D symptom capture, illness summaries, diagnosis and CPT codes, and medication and lab-order suggestions.
I built document classification and extraction pipelines, worked on billing automation, and led junior data scientists. Related work covered clinical coding and claim-denial prediction.
The document pipeline ran in a healthcare environment with PHI-aware logging. Data handling was part of the system design.
The engineering decisions
Design for inconsistent input. The pipeline processes scanned and handwritten records rather than assuming clean digital text.
Connect prediction to the workflow. Classification and extraction feed structured data into the billing process. The useful outcome is less manual work and faster processing.
Build the data foundation. SQL Server, SSIS, reporting models, and embedded analytics made the outputs accessible to operational teams.
A model metric is useful when it connects to the process around it. Classification accuracy and claim-cycle time measure different parts of the same delivery problem.
How the work was measured
Document classification achieved 98%+ accuracy. Operational metrics tracked manual work and claim-cycle time.
Related clinical NLP work reached micro-F1 0.81 on ICD-10 coding. A claim-denial model delivered 2× lift over rules. Those figures describe separate models, not the document classifier.
What changed
The production pipeline processed 60K+ healthcare documents per day. Document and billing automation reduced manual work 60%, while claim-cycle time fell from 24 hours to 6.
The work reinforced a principle that carries into my agent engineering: the model is one component in a larger operational system.
A public engineering summary. Results are scoped to the project and period described; implementation details are summarized at a high level.