MedFax AI

Inspiration

Physician offices still receive important referrals, pathology reports, and follow-up records by fax. The information is often scattered across long documents, hard to scan quickly, and sometimes incomplete or contradictory. We wanted to make that first review faster without pretending AI should replace clinical judgment.

What it does

MedFax AI is a protected physician fax workspace. It turns an incoming fax into a concise clinical brief:

• A short patient summary
• What information is still missing
• Evidence-backed draft recommendations with source citations

It first checks whether the documents belong to the same patient and whether key facts conflict. If the packet is mixed, incomplete, or explicitly unverified, it withholds clinical analysis and asks for physician reconciliation instead of guessing.

For supported cases, the interface can also show a simple anatomy orientation cue—for example, a prostate marker for a fax that explicitly concerns prostate cancer. It is an orientation aid only, not diagnostic imaging or a lesion map.

Every output is draft decision support. A physician must review the source fax and accept the draft; the app never places orders, diagnoses a patient, or sends a fax automatically.

How we built it

We built MedFax AI with Next.js, React, FastAPI, Supabase, and Telnyx fax workflows.

GPT-5.6 powers the server-side workflow: faithful fax extraction, document reconciliation, structured clinical summaries, missing-information detection, routing, and evidence-grounded draft review. We use structured outputs and keep OpenAI requests non-persistent with store: false.

Codex was our development partner throughout the project. We used it to design the two-service architecture, build the document-integrity safety gate, improve the physician interface, create the citation hover experience, add local verification, and test the production build.

Challenges we ran into

The biggest challenge was realizing that a clean-looking summary can still be unsafe if the source packet is inconsistent. In one synthetic packet, records contained conflicting disease history and even an explicitly synthetic molecular report. A normal summarizer could blend those details together.

We changed the workflow so document reconciliation happens before any triage or recommendation. The system now identifies conflicts, tracks which pages can be used, and withholds downstream clinical content when it cannot safely establish a single patient narrative.

We also worked hard to make the interface concise. Physicians need the signal first—not a wall of generated text—so we reduced the default view to patient context, missing information, and evidence-backed recommendations, with details available on demand.

Accomplishments that we're proud of

• Built a real end-to-end fax workflow instead of a chat-only demo.
• Added a fail-closed document-integrity gate for mixed or conflicting records.
• Kept the physician in control at every important decision point.
• Added hoverable citations so a clinician can see where an evidence claim came from.
• Created a short, readable clinical brief rather than another long AI report.
• Added a small anatomy cue that makes the main issue easier to understand at a glance.
• Verified the web app with linting and production builds, plus local service checks.

What we learned

We learned that medical AI is often more useful when it knows when to stop. The most important feature was not generating more text—it was refusing to produce clinical recommendations when the documents could not be reconciled.

We also learned that trust comes from traceability. A concise claim with a visible source is much more useful than a polished paragraph with unclear provenance.

What's next for MedFax AI

Next, we want to expand the supported anatomy cues and specialty-neutral routing, add clinician feedback loops for improving summaries, build a formal evaluation set for extraction and reconciliation, and improve the referral follow-up workflow.

Most importantly, we would validate the system with clinicians before any real-world clinical deployment. MedFax AI is currently a hackathon research prototype, not a medical device or replacement for physician judgment.

Built With

• gpt5.6