MerlinPDF

Inspiration

Academics and researchers around the world are confronted with millions of papers of dense text. This is partially due to the paradigm of how papers are written, as well as the difficulty to get ideas across.

What it does

An advanced highlighting tool for academic journals with text to chart inference for improved document understanding. This tool will accelerate learning and and comprehensibility of complex, multi-domain papers.

How we built it

Frontend: React, Next.js, Plotly, REST API
Backend: FastAPI, OpenAI API using DigitalOcean, openai-gpt-oss-20B
Infra: Docker, Docker-compose

Challenges we ran into

PDFs are very challenging to handle, due to their compressed, non-transparent nature. Text is not naturally selectable, and most libraries for addressing this are deprecated and non-functional. Prompt fine-tuning to reduce false positives was a big challenge. This was addressed using few-shot prompting so that the model had example scenarios to judge its decision making off.

Accomplishments that we're proud of

We managed to compose a full-stack application with proper dev ways of working, by connecting a FastAPI backend and next.js frontend. We also created an AI agent on DigitalOcean. This agent converts unstructured dense text into a structured JSON schema to generate charts. We are very proud of this, as conceptually it was a challenge to figure out our ontology.

What we learned

The frontend can, somehow, suck all of your time :( Also, PDFs are challenging to work with. But that was the challenge we wanted to address, so no complaints from us :)