Inspiration

The inspiration for Universal Document Interpreter came from seeing how difficult and stressful it can be for people—especially migrants, travelers, and workers—to understand and manage important documents like passports, visas, and contracts. Missing a deadline or misunderstanding a clause can have serious consequences. We wanted to build a tool that makes document intelligence accessible, fast, and easy for everyone.

What it does

Universal Document Interpreter lets users upload one or two documents (PDF, DOCX, or images) and instantly receive:

Plain-language summaries of the document’s content
Risk analysis and urgency badges
Action items and next steps
Expiry timelines and visual progress bars
Downloadable PDF summaries
All with privacy-first, stateless processing

How we built it

Frontend: Built with Next.js, Tailwind CSS, and shadcn/ui for a modern, glassmorphic dashboard. Features include drag-and-drop upload (react-dropzone), skeleton loaders, copy-to-clipboard, and PDF download (jsPDF).
Backend: FastAPI (Python) handles file uploads, OCR conversion (Docling), PII redaction (Presidio), and AI analysis (Ollama running Llama3). Async background tasks ensure smooth UX, and multi-doc synthesis logic compares documents (e.g., passport + visa).
AI Models: Uses local LLMs via Ollama for analysis, and supports multiple OCR backends for flexibility and speed.

Challenges we ran into

Integrating multiple AI tools (OCR, PII redaction, LLM) into a seamless workflow
Handling large, multi-page documents efficiently and reliably
Ensuring privacy and statelessness while providing rich analysis
Designing a UI that feels modern, professional, and easy to use for all audiences
Managing asynchronous processing and real-time status updates between backend and frontend
Less time spent on backend inference

What we learned

How to orchestrate multiple AI and NLP tools for practical document intelligence
The importance of clear, actionable UX for users dealing with high-stakes documents
Best practices for privacy, stateless processing, and local inference
How to design and build glassmorphic, high-tech UIs with Next.js and Tailwind CSS
The value of rapid prototyping and iterative improvement in AI-powered products