Patient Co-Pilot: A Source-Grounded Companion App for Navigating Complex Care

Patients living with complex or longitudinal health conditions such as cancer often accumulate dozens to hundreds of clinical records (clinic notes, recordings of doctor visits, pathology and imaging reports, laboratory reports, and medication changes) across multiple providers and over time. Between visits, especially during the diagnostic process, patients face the difficult task of making sense of changes in their condition, interpreting updates about their disease, deciding when a new or worsening symptom may require urgent attention, and managing travel, work, and family life around a disease they did not choose to learn about. We present Patient Co-Pilot, a companion mobile application built around this patient experience. It was first developed for a friend navigating a complex disease and was subsequently extended through use by six additional TestFlight beta users. The app ingests a patient's own records, organizes them into a dated longitudinal history across four in-app surfaces (Voices, Docs, Reports, and Home), and answers questions using safety-gated, guardrailed pipelines with exact, tappable citations to the patient's own evidence. Seven safety gates check information at key stages, from importing and summarizing records to answering questions and preparing information for clinician review. Patients can review the AI-generated information and inspect its supporting evidence rather than relying on model confidence alone. The goal is not to reproduce physician decisions, but to help patients recognize important action windows in which a symptom, medication change, or major life decision may require attention, clarification, or communication with their care team, thereby giving them greater control over their treatment journey. In a preliminary evaluation on one patient's records, combining retrieval, inspectable source grounding, and post-generation verification reduced the hallucination-flag rate and increased faithfulness. Rather than demonstrating general model superiority, these preliminary results illustrate how existing models can be structured with retrieval, inspectable evidence, and verification for sensitive patient-facing applications. Reviewer resources: An anonymized demonstration video and reviewer code are provided in the supplementary material.