Inspiration
MD_ASSIST began with a simple, frustrating experience: at a first specialist visit, the physician had no current patient data and spent 20–30 minutes reconstructing medical history before the patient opened MyChart and supplied recent results. That was the “aha” moment. Valuable clinical time should be spent understanding the patient and planning care—not hunting for information that already exists.
What it does
MD_ASSIST is designed to turn scattered patient-provided health information into a concise, clinician-ready brief before the visit. It helps surface the essential context a clinician needs so the conversation can begin from a shared understanding rather than from a blank page.
The project supports preparation and communication; it does not replace a clinician, diagnose disease, or prescribe treatment.
How we built it
MD_ASSIST was developed as a solo project with Codex as the primary development environment and GPT-5.6 at the center of the AI workflow. Codex supported rapid iteration across the product workflow, implementation, testing, and documentation.
To protect privacy during development and demonstration, we used synthetic patients and synthetic medical records rather than real patient data. We also created documented evaluations to test the system’s behavior and the usefulness of its outputs.
Challenges
The hardest challenge was balancing compression with clinical usefulness. A short brief can save time only if it preserves the facts that matter, distinguishes known information from uncertainty, and avoids presenting generated text as medical judgment.
A second challenge was treating safety and privacy as product requirements from the beginning—not as features to add later.
Accomplishments
- Built a working end-to-end prototype as a solo developer
- Produced a complete demonstration video
- Used synthetic records and patients throughout the demo workflow
- Created documented evaluations for repeatable testing
- Established a clear boundary between information preparation and clinical decision-making
What we learned
The biggest lesson is that better clinical encounters do not always require more data. They require the right information, organized at the right moment, with clear provenance and responsible human oversight.
Codex made it possible to move from a personal insight to a tested prototype quickly while keeping the work structured and reviewable.
What’s next
Next steps include expanding evaluation coverage, strengthening source traceability, refining clinician-facing summaries, and validating the workflow with qualified healthcare professionals. Privacy, security, accessibility, and regulatory review will remain central as MD_ASSIST develops.
Built With
- cloudflare
- codex
- gpt-5.6
- next.js
- react
- tailwind-css
- typescript
- vite
Log in or sign up for Devpost to join the conversation.