ChatPDF — AI-Powered PDF Assistant
Backend Developer · Graduation Project
The problem
A student with a 140-page PDF has no way to ask it a question. Reading it is the only interface. The project set out to make a document answerable — in chat, in summary, as a quiz — and to do it for scanned and handwritten pages too, not just clean digital text.
Constraints
- No budget for hosted inference, so every model had to run on hardware the team had.
- Answers had to be traceable to the source document rather than invented, which rules out a bare language model.
- Six people, one semester, and an academic defence at the end of it.
Key decisions
Retrieval-augmented generation chosen over fine-tuning alone
A fine-tuned model knows what it was trained on; it cannot answer about the PDF a user uploaded five minutes ago. Retrieval puts the actual document in front of the model at question time, which is also what makes an answer checkable against a source.
QLoRA on a 7B model chosen over full fine-tuning
Four-bit quantisation and low-rank adapters brought training within reach of the GPUs actually available. Eight configurations were compared; Mistral-7B scored highest.
A small CNN trained from scratch chosen over VGG19, ResNet50, Xception and EfficientNetV2
Seventeen architectures were trained and measured for the handwriting recogniser. The large pretrained networks all did worse here than a purpose-built three-layer CNN — Xception reached 0.34 training accuracy where the small model reached 0.88 on test.
An AI-powered PDF application that revolutionises document management, built as our graduation project at Arab International University (team of 6).
Features: chat with your PDF using LLMs and Retrieval-Augmented Generation (RAG); summarization of extensive documents; OCR for handwritten images; interactive quiz generation from PDF content (MQAG framework); audio-to-text transcription of lectures using Whisper; PDF editing, merging, splitting, and format conversion; user profiles and an admin dashboard.
I built the backend infrastructure for both the mobile application and the admin dashboard.
Outcome
- Retrieval found the correct source passage in the top three results for 96 of 100 test questions
- Best fine-tune (Mistral-7B) reached 0.95 BERTScore F1 and 0.64 BLEU across eight compared configurations
- Handwriting recogniser reached 0.88 test accuracy, chosen after training seventeen architectures
- Defended at Arab International University in February 2024
Technologies
University project, built during my B.Sc. at Arab International University.
Want something like this for your business?
Let's talk