Transaction Coding Agent
Bank lines in → suggested accounts, tax codes, and confidence → accountant Approve / Edit.
Suggested coding with confidence scores, review queue, and activity log.
30-second recap
- Bank lines in → suggested accounts, tax codes, and confidence → accountant Approve / Edit.
- Suggested coding with confidence scores, review queue, and activity log.
Transaction Coding Agent
The interactive console loads on demand. Upload your own file or edit the example input. Every result comes back live from the API.
The problem
Bookkeepers spend hours categorizing bank transactions with inconsistent rules and little traceability.
How it works
Upload or paste transactions → agent drafts coding against a chart of accounts → reviewer approves, edits, or escalates → decisions land in an audit trail.
System design
Model-agnostic classification agent via provider-agnostic LLM API with tools for CoA lookup and rule hints. Structured output: account, tax, confidence, short rationale.
Human review & control
Rows below a confidence threshold queue for review. Every action records actor, timestamp, before/after values.
Integration scope
Mock ledger adapter (Xero / QBO-shaped). Swap-ready for real OAuth later (mock).
Stack
- LLM agents
- Next.js
- TypeScript
What the live demo shows
Live structured LLM output with account, tax, confidence, and rationale. Review queue and audit log run in the app.
Current limitations
Ledger destination is mock. No real Xero or QBO connection, no OAuth, no posting.