Reconciliation Workbench is an audit "tie-out" desk on a live financial statement. A small model, MiniCPM-V 4.6 (~1.3B), reads the report, while a deterministic engine does every piece of arithmetic to catch the numeric errors OCR introduces: leading-digit insertions, trailing truncations, column misalignment and garbled characters. Every number traces back to its place in the source PDF and explains its reconciliation relationship; in the full version it can also be hand-corrected and re-checked, leaving an exportable audit trail.

It is a Build Small Hackathon submission on the Backyard AI track.

The problem

This is the unglamorous, real-world problem auditors face every day: a 300-page Chinese annual report or prospectus is OCR'd, and a single misread digit silently breaks a balance sheet. Reconciliation Workbench catches those errors and never "helpfully" rewrites the source.

Why small

You do not need a giant LLM to audit a financial statement.

• Reading is for a small model: MiniCPM-V 4.6 has ~1.3B parameters, well under the Tiny Titan 4B bar.
• Arithmetic is for code: every add, subtotal and cross-foot happens in a pure-function deterministic engine.

Moving the error-prone math out of the model is exactly why it catches OCR errors instead of inventing new ones: a model that both reads and computes could hallucinate a plausible-but-wrong total.

Try it

The Space bundles two samples and needs no keys:

1. Sanmu Group 2024 Annual Report: a real PDF with per-digit pixel highlighting and a tri-state overlay (green = reconciled, red = suspected OCR error, grey = not covered by rules).
2. A prospectus: the engine catches subtotal reconciliation errors on the consolidated balance sheet, with a verification queue you can step through.

Pick a sample in the top-right corner and open each finding; click a flagged number to see it light up in the PDF, the live sheet and the focus card at once.

This is a read-only public demo. Live PDF upload and the human-correction → recompute → audit-trail loop run on the deterministic engine, which is proprietary and not shipped in the public demo. The demo video below shows those features in action.

How it works

Plain Text

A single custom gr.HTML island (pdf.js + live sheet + verification panel + all interaction)
  └─ payload: small model reads + structures → deterministic engine reconciles → tri-state findings

• Custom UI: not a stack of default Gradio components, but one custom gr.HTML island with server_functions, a real pdf.js viewer, tri-state bbox overlays and a keyboard-driven review queue.
• Iron rule: extraction and arithmetic are separated; the raw OCR value is immutable; a human correction is a new, traceable transaction, and the original is always kept as a reference column.
• Fail loud: the UI reports "N tables dropped during location · 0 findings ≠ error-free" and never disguises "couldn't extract" as "nothing wrong".

Demo video

The 3-minute demo video walks through the full loop: the 1.3B model reads a real annual-report PDF → the deterministic engine flags a balance-sheet subtotal that doesn't tie → the engine-inferred correction is applied → it recomputes and re-ties (✓ Closed), with a full, tamper-proof audit trail.

Limitations

• The public demo ships only the frontend and pre-computed sample results.
• The deterministic reconciliation engine and the extraction pipeline are proprietary and not included, so the public version does not support uploading your own PDFs or the human-correction loop.

Links

• Hugging Face Space: build-small-hackathon/reconciliation-workbench
• Demo video: demo.mp4
• Social post: X

Built by niuniu869.