UPIDigest

Problem

While planning a personal trip, I realized there was no reliable software that could cleanly analyze Google Pay (UPI) statements alongside traditional bank statements. Existing tools were fragmented, inaccurate, or required heavy manual categorization. UPIDigest was built to solve this exact problem end-to-end.

System Overview

• Built a web application to ingest UPI transaction statements and bank statements from multiple sources
• Designed reconciliation logic to automatically match overlapping transactions across UPI and bank records
• Normalized raw statement formats into a unified transaction schema for consistent analysis
• Classified transactions into spending categories and identified recurring patterns

Data Engineering & Matching Logic

• Implemented deterministic matching using amount, timestamp windows, merchant identifiers, and transaction IDs
• Handled edge cases such as partial matches, delayed settlements, and duplicate entries
• Maintained a clean separation between raw ingested data and derived analytical views
• Designed the pipeline to be repeatable and idempotent across multiple statement uploads

AI-Powered Expense Digest

• Integrated an on-device LLM using Ollama with Mistral 7B for local, privacy-preserving inference
• Generated a daily natural-language summary highlighting spending behavior, category-wise breakdowns, and anomalies
• Ensured all summaries were grounded strictly in computed transaction data, avoiding hallucinated insights
• Designed prompts to produce concise, human-readable financial narratives instead of raw statistics