BillSense, built by Team Demomorgans for Problem 3B: Extract Information from Receipts, is an AI-powered web application that converts receipt images into structured expense records. Users upload a receipt, extract its company name, date, address and total, review or correct the information, and save it for expense organisation.

We specialised the existing Gemma 4 vision model for receipt extraction using Low-Rank Adaptation (LoRA), a way to fine-tune an AI model without changing all its original parameters; through Unsloth. This approach trains a small set of adapter parameters on labelled receipt examples while keeping the base model’s weights frozen, reducing the resources needed compared with full-model fine-tuning. Our training pipeline supports ICDAR2019-SROIE receipts, synthetic Indian bills including handwritten bills and custom labelled images, alongside an evaluation pipeline for comparing original and fine-tuned model outputs with the help of LoRA.

Built with React, TypeScript, FastAPI and SQLite, BillSense connects receipt extraction with practical features including organised/unorganised bill classification, saved bill folders, bill comparison and CSV export. It brings receipt processing and expense organisation into one workflow, with user review before saving.