Chart Pilot

ChartPilot takes a recorded conversation between a patient and a clinician and turns it into documentation the clinician can trust. After consent is recorded, the audio is transcribed and the clinician reviews and corrects the transcript. ChartPilot then extracts clinical facts, drafts a SOAP note (Subjective, Objective, Assessment, Plan) in which every statement links back to the part of the conversation it came from, suggests ICD-10-CM diagnosis codes, and produces a FHIR R4 document that the clinician previews, confirms and downloads.

It is built for clinicians who spend hours on documentation after visits, and for the practices and health systems that want to give them that time back. It also tackles the problems that companies building AI scribes and ambient documentation tools have to solve. A clinician moves through one clear workflow (consent, audio, transcript review, note review, coding review, export) with a dashboard and timeline showing where each visit stands.

The project focuses on the engineering around the AI model, which is where documentation tools usually fall short. Nothing advances without a clinician's approval: the transcript must be approved before a note is generated, the note before codes are suggested, and the export before it is locked. Every edit creates a new version instead of overwriting the old one. Transcripts are treated as untrusted input, so instructions hidden inside a conversation (prompt injection) are ignored, and one of the demo scenarios tests exactly that. A safety layer blocks inferred facts, invented doses and unsupported statements, and diagnosis codes can only come from a validated list checked against the official CDC and CMS releases.

Every record is scoped to an organization, audit logs never contain transcript text, and audio files are stored privately and served only through access-checked routes. ChartPilot never creates orders or prescriptions from AI output, never bills, and never sends data to outside systems. The demo uses predictable stand-in transcription and language-model providers so it runs reliably without cloud accounts, while adapters for AWS Transcribe Medical, Bedrock and Cognito are built in for a cloud deployment. The July 2026 release check passed 80 backend tests, 7 front-end component tests, linting, type checks, production builds and the full upload-to-confirmed-export workflow. It uses fictional data only and makes no HIPAA-compliance claim.