It started with my granddad's checkups.
My granddad gets regular checkups, and each one leaves another report full of numbers. I wanted a way to keep track of them all and actually understand what they meant, and whether anything was changing from one visit to the next.
MedIntel reads a report (a PDF, a scan or a phone photo), pulls out every finding with the exact text it came from, summarises it, flags what looks worrying and why, and tracks it across reports. You can then ask questions in plain language and get answers that cite their source.
From upload to answer.
Every patient and report here is fictional demo data; no real medical data is involved.
Every report walks the same path.
Each upload moves through a fixed sequence of stages. Every stage can be re-run without duplicating anything, and if one fails, the original file and its text are always kept.
Medical numbers leave no room for guessing.
Risk comes from a fixed, weighted rule set over the extracted findings, so the same report always gets the same score. Every score is labelled as screening help, not a diagnosis.
A guardrail checks every value and name in a summary against the report. Anything it can't trace is stored as an issue, not quietly shown.
Chat answers come from retrieved passages, with page and section citations. If nothing relevant is found, it says so instead of guessing.
Extraction, summaries, risk and search all have deterministic fallbacks. Only free-form chat needs the language model.
Each risk factor links to the finding behind it, and each finding links to the highlighted text in the original report.
Summary, risk and findings, exported as one document.
Measured against hand-checked reports.
An evaluation script runs the system against nine synthetic reports written to be awkward: missing values, repeated readings, multi-page layouts, poor formatting. Each metric has a pass mark set in advance.
| Metric | Result | Pass mark |
|---|---|---|
| Extraction: value, unit and status | 100% (41 findings) | ≥ 90% · pass |
| Summary grounding (deterministic) | 100% (9 summaries) | ≥ 95% · pass |
| Summary grounding (stored, incl. LLM) | 92.3% (13 summaries) | reported |
| Retrieval: right passage in top 1 / 3 / 5 | 100% / 100% / 100% (27 questions) | top-5 ≥ 90% · pass |
| Risk checks: score, level, factors, confidence | 13 / 13 on each | all pass |
These are synthetic test reports, not a clinical validation.
Five layers, one command to run.
A React dashboard talks to a FastAPI backend, which separates the API, the business logic and the AI components, all on PostgreSQL with pgvector. docker compose up brings up the whole stack with the demo patients already loaded.