ClearPath

Inspiration

Every day, millions of people across the EU receive medical letters, insurance documents, and legal notices filled with jargon they cannot understand. For older adults, non-native speakers, and anyone without a medical or legal background, a single paragraph can mean the difference between following treatment correctly and missing a critical deadline.

We were inspired by public-service kiosks the kind you see in hospitals and government offices where clarity and accessibility matter more than flashy design. We asked: what if understanding your own paperwork was as simple as using an ATM?

What it does

ClearPath is a kiosk-style web application that:

Accepts - a medical or legal document (PDF or plain text)
Extracts the text and sends it to Groq AI
Returns a structured, plain-English analysis with:
- Summary -- 2 sentences explained like you're 10 years old (ELI5)
- Action Items -- 3 concrete steps to take (with checkboxes)
- Glossary -- difficult terms with simple definitions The app is fully in English, designed for the EU market, and built as a navigable multi-panel application (Home, Upload, Analysis, Results, History, Help) not a marketing landing page.

How we built it

Frontend: Next.js 16 (App Router), React 19, TypeScript, Tailwind CSS v4, shadcn/ui, Plus Jakarta Sans
AI: Groq API (openai/gpt-oss-120b) via Vercel AI SDK with streamObject for real-time structured JSON output
PDF parsing: unpdf for server-side text extraction
Schema validation: Zod schema enforcing summary, action_items[3], difficult_words[]
Deployment: Vercel with GROQ_API_KEY as environment variable
Development: Cursor + Claude for rapid iteration

Architecture flow:

Upload PDF → /api/extract (unpdf) → plain text → /api/analyze (Groq streamObject) → UI panels

The kiosk UI was designed with sharp panels, high-contrast teal accents, and large touch targets optimized for clarity over aesthetics.

Challenges we ran into

PDF extraction vs. demo shortcuts: Our initial hackathon MVP sent hardcoded sample text instead of reading the uploaded file. We caught this during testing and implemented real PDF text extraction with unpdf.
Scanned PDFs: Image-based (scanned) PDFs cannot be parsed without OCR, we added clear error messages and documented supported formats.
Model deprecation: Groq deprecated older Llama 3 models during development; we migrated to openai/gpt-oss-120b for structured output support.
From landing page to application: Our first version looked like generic AI slop. We redesigned into a kiosk-style multi-panel app with proper navigation and institutional UX.

Accomplishments that we're proud of

Built a working end-to-end pipeline real PDF upload → text extraction → AI analysis → structured UI in under 24 hours
Achieved near-instant results thanks to Groq's LPU inference speed
Designed a genuinely usable kiosk interface with Home, History, Help, and step-based workflow
Zero cost to run on Groq free tier + Vercel free tier
Structured AI output that is actionable, not just a wall of text

What we learned

Structured output (streamObject + Zod) is far more useful than free-form chat for document analysis users need predictable sections, not paragraphs
Speed matters: when analysis takes under a second, the product feels trustworthy; when it takes 5+ seconds, users assume it broke
Kiosk UX patterns (step indicator, sidebar nav, action bar) work better than scrollable landing pages for utility apps
Always test with a real uploaded file demo shortcuts hide broken pipelines

What's next for ClearPath

OCR support for scanned PDFs and image uploads (JPG, PNG)
Multi-language output: English, Polish, German, French for broader EU coverage
Document type detection: auto-tune prompts for medical vs. legal vs. insurance documents
Printable action plan: export summary + checklist as PDF
Accessibility audit: WCAG 2.1 AA compliance for public kiosk deployment
Persistent history with optional anonymous session storage