Financial audit with local AI

Auditing a large body of documents without data going outside: conversion, anonymisation, RAG analysis on local models

A project to detect misuse of a corporate group’s funds: a body of documents (about 3.9 GB in Markdown) is processed locally or on secure servers. The work produced a study of LLMs and GPU infrastructure, a final “quality vs. time” scenario on a single H200 GPU with Qwen2.5-72B and two-tier escalation (a small model picks out suspicious items, a large one does the detailed analysis), a cost estimate and a roadmap. Tools were built for conversion to Markdown, reversible anonymisation of personal data and RAG analysis on Milvus; a trial analysis was carried out with a final report.

The task

To establish where, when and why the money was spent, without passing confidential data to external services.

What’s inside

• Conversion of PDF, DOCX, XLSX and other formats to Markdown
• Reversible anonymisation (Presidio, spaCy NER, Faker): full names, organisations, phone numbers, emails, dates, amounts, INN, SNILS, OGRN, bank accounts, cards, IBAN; realistic substitution and placeholder modes, date shifting, amount scaling, a mapping table for de-anonymisation
• RAG analysis of related documents (contracts, invoices, payment orders, letters, completion certificates) and a graph of entities and transactions with anomaly detection rules
• An air-gapped mode for working without external APIs; reports in Markdown, JSON and Excel

Machine translation — being proofread.