01 / PROJECTS / FINDOCPARSER

FinDoc
Parser.

Take a photo of a receipt, invoice, or bank statement and it pulls out the important details for you: who you paid, how much, and when. Before it trusts an answer, it checks its own work several different ways and tells you how confident it is.

Most document AI stops at extraction. The interesting engineering here is the trust layer: five independent signals that vote before anything is reported to the caller. Clean results auto-accept. Shaky ones route to a human reviewer.

Index / Interactive demo

Simulation only
FinDocParser — pipeline simulation

Select a document

Pipeline

  1. Difficulty routing
  2. OCR coverage
  3. Cross-model agreement
  4. Math check
  5. Agentic review

Select a document and press Run to simulate the pipeline.

This is a front-end simulation of the real pipeline. No files leave your browser.

Index / How it actually works

Extraction is the easy half. A multimodal model reads the page image and returns structured fields in a single pass. The hard half is deciding whether to trust the result. That is done with five independent signals that vote before anything is reported.

  1. 01

    Difficulty routing

    Before extraction starts, a lightweight pass grades how hard the document is and sends it to the right model tier. A clear printed receipt goes to a cheaper, faster model. A handwritten or scanned form goes to a more capable one.

  2. 02

    OCR coverage check

    A classical OCR pass runs in parallel. Every extracted value is checked against the raw character data to confirm the numbers actually appear on the page, not just in the model's imagination.

  3. 03

    Cross-model agreement

    For fields the primary model is uncertain about, a second model independently reads the same region. If they disagree, neither answer is trusted automatically.

  4. 04

    Deterministic math

    Line items are summed and checked against the stated total. This is an independent signal that cannot hallucinate: the arithmetic is either right or it is not.

  5. 05

    Agentic review

    A final-pass agent reads the assembled result and sanity-checks the whole thing: does the date look plausible, does the vendor name appear anywhere on the page, are there contradictions between fields?

The output carries a confidence score that aggregates all five signals. Scores above a threshold auto-accept. Scores below route to a human review queue with a structured explanation of which signals failed and why. The system knows when not to trust itself.

Interested in document intelligence for your stack?

get in touch