✦ Audit SaaS Platform with an AI Mapping Agent

The problem

Audit teams ran everything in Excel. Every client’s chart of accounts looked different, so mapping ledgers to audit standards was slow, manual and error-prone — and every change in accounting standards meant editing formulas across hundreds of workbooks.

What I built

I co-founded the company and led a five-person team that built an audit SaaS platform from scratch:

• One microservice per audit subject (cash, receivables and payables, inventory, revenue …), each with its own repository, database schema and deployment, on top of shared platform services: a workpaper engine, cross-check rules and multi-tenant access control.
• Rules as data. Formula logic moved into a versioned DSL, so a new accounting standard is a configuration change, not a code change.
• An AI Mapping Agent built with LangGraph and FastAPI that maps each client’s ledger accounts to the audit standard. High-confidence matches are applied automatically; low-confidence ones go to an auditor, whose corrections feed back into the agent.

← swipe to see the whole diagram →

Key design decisions

1. 1One service per subject — own repository, database and deployment, so subjects never break each other.
2. 2Rules as data — formulas live in a versioned DSL; a new accounting standard is a config change.
3. 3Confidence triage — confident mappings fill workpapers automatically; uncertain ones go to an auditor.
4. 4Feedback learning — auditor corrections are written back to Milvus, so the next mapping is better.

Left: auditors use web, desktop and admin clients through an API gateway to one microservice per audit subject, which use a shared platform (cross-check rules, workpaper engine, versioned rule DSL, row-level security). Right: client ERP exports go through ETL and validation to an AI mapping agent that retrieves similar mappings from Milvus; high-confidence results are written into workpapers automatically, low-confidence ones go to an auditor, and the corrections are written back to Milvus. Subject services trigger mapping jobs through Kafka events.

The impact

• 70% of manual SOP workflows automated, cutting overall manual work by about half.
• Accounting-standard updates now need only a configuration change.
• Load-tested for 50K users at 10K peak QPS.
• Handed over to the operating team in February 2025.

Key decisions

• Triage, not full automation. In audit, a confident wrong answer is worse than no answer, so the agent only acts on its own when it is sure.
• Isolate each audit subject. Separate services let each subject evolve with its own rules without risking the others.
• Make rules versioned data. Auditors can see exactly which version of a rule produced a number.