Inspiration
A relative of mine came home from hospital with a discharge note nobody in the family could decode. Dose timings, warning signs, follow-up dates, all buried in one dense page. The nurse had explained it once, fast, at midnight. We guessed.
That is normal, and that is the problem. Discharge notes, drug labels, and insurance forms are written at college level, while millions of readers are left to guess: people with dyslexia, ADHD, or low vision, elderly patients, low-literacy readers, anyone reading in a second language. A misunderstood dose can send someone straight back to hospital. I had been working on readability tooling, and healthcare was the place where it clearly mattered most. That became MediClear.
> MediClear turns confusing prescriptions, discharge notes and health pages into plain-language cards, checklists and listen mode for dyslexia, ADHD, low-vision and elderly patients.
What it does
Paste a health link or page text. MediClear returns:
Plain cards. One idea per card, short words, short sentences. Never a wall of text.
A numbered checklist. Dose steps, appointment dates, and prep instructions pulled out as "Step 2 of 4" with urgency flags.
Listen mode. Reads the page aloud with word-by-word highlighting for low-vision readers.
Ask this page. Questions answered using only the page text. The original stays visible beside the output, so families can verify nothing was dropped or invented.
A caregiver sheet. One clean printout of the simplified page.
Scope is deliberate: MediClear explains existing instructions. It never diagnoses, never prescribes, and cites the source on every page.
How we built it
This round is the idea pitch, so this section is the build plan: three stages, all standard tooling, no new research required.
Clean. Page HTML is stripped of ads and navigation with an open-source readability extractor. No AI involved.
Restructure. A language model reshapes the cleaned text into strict JSON: title, summary, action items, and sections. Malformed output is retried once, then surfaced as a clean error. Never a crash.
Render. Cards, checklist, listen, and Q&A are simple renderers over that same JSON. Planned stack: Next.js, React, TypeScript, serverless hosting. No database, no login.
Success is measurable. Readability grade follows the Flesch-Kincaid form:
$$Grade = 0.39 \times \frac{words}{sentences} + 11.8 \times \frac{syllables}{words} - 15.59$$
The pilot target takes sample pages from $Grade_{original} \approx 14$ (college) down to $Grade_{simple} \le 5$, with patients actually finishing their checklists:
$$Completion = \frac{steps\ done}{steps\ assigned} \ge 0.8$$
Challenges we ran into
Invented medical facts. The hardest risk, and the first one I designed against. Answer: grounded generation only, original text pinned beside every output, and a no-diagnosis boundary written into the product.
Sites that block page fetching. Hospital portals do this. Answer: a paste-text path from day one, plus pre-saved sample pages so demos never depend on live sites.
One design cannot fit all. A dyslexic reader and a low-vision elderly patient need different things. Answer: separate renderers (dyslexia font and tint, ADHD micro-cards, listen highlight) over one shared JSON, instead of one compromised view.
Fitting a health idea into 6 slides. The template forces hard choices. Answer: one claim per slide, every claim traceable to a named reference, numbers cut unless verifiable.
Accomplishments that we're proud of
A concept with a sharp boundary: explains, never diagnoses. Most health AI pitches blur this. Ours does not.
A pipeline with no research risk: every stage maps to existing open tooling, so the Final Round build is assembly and design, not invention.
A measurement plan, not vibes: grade drop plus checklist completion as the two numbers that decide if it works.
A complete idea-round package: 6-slide PDF in the official template, matching Devpost story, built solo in three days.
What we learned
Splitting content from presentation was the key insight: restructure once into clean data, then render many views. It makes every new reading mode cheap and keeps the AI on a short leash. The second lesson: in health, restraint is a feature. The most important thing MediClear does is what it refuses to do.
What's next for MediClear
Build the health renderers (dose-first cards, appointment checklists, caregiver print sheet), assemble a cached demo set from public patient-education pages, and pilot with a university clinic or local pharmacy tracking grade drop and checklist completion. If shortlisted, the Final Round gets the working proof. Long term: prescriptions photographed from paper labels, and reading-level tiers tuned per condition.
Built With
- accessibility
- ai
- api
- digital-health
- llm
- nextjs
- openrouter
- react
- readability
- typescript
- vercel
Log in or sign up for Devpost to join the conversation.