Healthcare AI
TheraCare — AI-Powered Post-Discharge Follow-Up Agent
One in five Medicare patients is readmitted within 30 days of discharge, and over half of those readmissions are preventable. TheraCare parses discharge PDFs with an LLM, rewrites clinical text to a 6th-grade reading level, runs a structured daily patient check-in with validated PHQ-2 and GAD-2 screeners, and computes a transparent 0-100 readmission risk score surfaced on a real-time clinician dashboard.
Problem Statement
1 in 5 Medicare patients is readmitted within 30 days of discharge, costing the US healthcare system roughly $26 billion a year, and over 50% of those readmissions are preventable. Patients go home with ten-page clinical documents they cannot read, no one follows up, and warning signs get missed until they become emergencies.
Tech Stack
React NativeTypeScriptNode.jsExpressSQLiteOpenRouter GPT-4.1-miniFHIR R4PythonREST APIsExpo
Key Contributions
- Led backend architecture and AI integration for a 6-person team (The Machiavellians) at Illinois Institute of Technology
- Designed a prompt-engineering pipeline for schema-enforced extraction of diagnoses, medications, appointments, and warning signs from arbitrary hospital discharge PDFs
- Built a plain-language rewriter converting clinical text to a 6th-grade reading level, an intervention shown in clinical literature to reduce readmissions by up to 50%
- Built a transparent rules-based readmission risk engine (0-100) grounded in the LACE index and HOSPITAL score, with every point documented and auditable
- Implemented validated PHQ-2 and GAD-2 mental health screeners directly into a 10-question daily patient check-in flow
- Architected a FHIR R4 data layer (Patient, Condition, MedicationRequest) for live Epic and Cerner EHR integration
- Built the SQLite persistence layer with a 5-table schema and 15 REST endpoints, plus a real-time clinician dashboard sorting patients by risk with clinical flags surfaced immediately
Results
- 15 live API endpoints covering the full clinical workflow, from discharge ingestion to risk surfacing
- Transparent 0-100 risk score computed per patient from clinical, demographic, and behavioral inputs
- FHIR R4 architecture ready for Epic and Cerner integration without code changes
- Recognized by Leap of Faith (LOF) as a top proposal for summer development funding
Engineering Decisions
- Chose a transparent rules-based risk engine over an ML model, because in a clinical setting a score a clinician can audit line by line is worth more than a marginally better score they cannot interrogate
- Enforced a strict output schema on the LLM extraction step so downstream code never has to defensively parse free-form model output
- Grounded the scoring logic in published clinical frameworks (LACE, HOSPITAL) rather than inventing a bespoke heuristic, so the score is defensible to clinicians
- Built the clinician dashboard as a single-file HTML/CSS/JS view served directly from Express, keeping the clinical surface dependency-free
Challenges Faced
- Discharge documents have no standard format, so the extraction prompt had to be robust across wildly varying layouts while still producing schema-valid output
- Balancing model-driven flexibility against clinical safety meant deliberately constraining where the LLM was allowed to make decisions: extraction and rewriting, never risk scoring
- Designing the FHIR resource mapping to be genuinely EHR-ready, rather than FHIR-shaped in name only, required aligning the internal schema to R4 resource definitions from the start