CareCoordinator

Patients and caregivers are often overwhelmed by scattered medical records, from discharge summaries and prescriptions to lab reports and clinical notes, making it hard to prepare for appointments or track a health history. General-purpose AI tools can introduce unsupported facts, while cloud services may expose sensitive health data to external APIs. CareCoordinator addresses both challenges through local inference and verification-first processing. It runs on an AMD Instinct MI300X GPU (192 GB) using ROCm and vLLM to serve Gemma-4-31B-IT across six specialized agents with continuous batching. When a user uploads documents and describes a goal, a Main Agent creates a structured Care Contract and delegates tasks to four parallel Worker Agents: Intake, Document Classifier, Record Summarizer, and Medication Checker. A Manager/Compiler Agent combines their findings into four sections: a plain-language summary, questions for the doctor, a medication checklist, and a clinician review section. The key differentiator is that verification is enforced through deterministic Python checks rather than left to the model. The backend validates inline citations against the original source text, while claims it cannot ground are marked “cannot confirm” and routed to clinician review instead of being presented as fact. A multi-layer safety gate halts normal processing when a configured emergency indicator is triggered, while historical risks are flagged for review. Users must approve a summary before it becomes final, and the “Improve” feature routes feedback to the relevant worker for revision and re-verification. In a demo with synthetic patient records, CareCoordinator generated a cited summary at a 7.5 reading grade level in 83 seconds, with 52 statements checked against the sources. It helps patients and caregivers organize their records and prepare for clinical conversations without diagnosing conditions or giving medical advice.