Case file · AI / Finance
EPS Dashboard is an AI bookkeeping tool that ingests financial documents and turns them into structured figures — with one rule: every number on screen is traceable back to the source file it came from. Drop in an invoice or a statement, the system extracts the line items, and each figure carries a link to the exact document and place it was read from. Nothing in the ledger is unattributed.
Extraction is the easy demo; provenance is the product. Bookkeeping only earns trust if someone can audit any figure back to paper, so the link between number and source is treated as a first-class part of the data, not an afterthought.
Why it exists
AI is good at reading documents and bad at being held accountable for what it read. Bookkeeping is the domain where that gap is unacceptable — a wrong figure is not a typo, it is a misstatement. EPS exists to test whether an AI extraction pipeline can be made auditable enough that a human can trust it without re-keying everything by hand.
Shape of the system
┌──────────────┐
│ documents │ invoices, statements, receipts
└──────┬───────┘
▼
┌──────────────┐
│ extract │ line items + amounts pulled out
│ (vision LLM)│
└──────┬───────┘
▼
┌──────────────┐
│ ledger │ every figure keeps a link back
│ + provenance│ to its source file + location
└──────┬───────┘
▼
┌──────────────┐
│ dashboard │ click any number → see the paper
└──────────────┘
Three hard lessons
1. Provenance has to be designed in, not bolted on. Threading a source reference through every figure from extraction to display is far harder retrofitted. The data model has to carry it from the first step.
2. Documents are gloriously inconsistent. Real financial paperwork comes in every layout, scan quality and language. The long tail of weird documents is where most of the engineering hides.
3. Traceability is what makes the AI usable at all. An auditable wrong number can be caught and fixed in seconds; an unattributed right number still has to be re-checked by hand. The link is the trust.
Stack