The problem
Accountants and small business owners spend a surprising amount of time hunting through spreadsheets and documents to answer basic questions. How much did we spend on software last quarter. What is the outstanding balance on invoice 0047. What were our total expenses in March. The data exists somewhere, but it lives in formats you have to search through manually, not formats you can ask questions to.
The approach
- Built a retrieval-augmented generation system that ingests financial documents — invoices, expense reports, transaction histories, and bank statements — and makes all of it queryable in plain English
- The ingestion pipeline chunks and embeds each document with metadata tagging so retrieval stays accurate across different query types, not just simple keyword matches
- The LLM layer sits on top to generate clean, readable answers and includes source citations on every response so users can always verify where the number came from
- Multi-model routing handles different query complexities — simple lookups to a fast model, nuanced financial analysis to a stronger one — keeping cost and speed balanced across the whole system
The result
Financial data lookups that used to take five to ten minutes of digging through documents now happen in seconds. Users can query in plain English and get accurate answers with the source document cited. Accurate across invoice lookups, expense categorization, and summary-level reporting across multiple time periods.