Skip to content
All work

Accounting / bookkeeping

A back office that cut 650 held bookkeeping jobs to 24

Bookkeeping, check writing with owner limits, and an owner Q&A that works on paper, by voice, or on a tablet.

Client
A 15,000-acre New Mexico land and hospitality operator
System
Ledger
My role
Design, build, and run
When
Sep to Oct 2026
Office overview, sample data
Sample dataMoney in, money out and the decisions waiting on the owner, on one page.

01 / Problem

What wasn't working.

Several business lines, years of paper, and an owner who answers questions on paper. Transactions piled up waiting on someone who knew what they were. Escrow history existed only as scanned pages.

02 / Approach

What I built.

  1. 01

    About 9,300 financial rows this year in one ledger, with AI-proposed classifications that a person approves.

  2. 02

    Owner Q&A: questions go out on a tablet or on printed packets. Answers come back by voice, handwriting, or a scanned page matched by QR code, and get applied after review.

  3. 03

    Check writing with owner-set limits and PIN approval.

  4. 04

    2025 escrow history: 1,970 rows read from 66 scanned pages, each page read twice by two different models.

  5. 05

    The home screen is a to-do list, not a dashboard.

Questions for the owner, sample data
Sample dataOpen bookkeeping questions written in plain language for the owner to answer on a tablet or paper.
Bank check register, sample data
Sample dataEvery bank check with its purpose, status and linked receipt.
Receipt intake, sample data
Sample dataFive ways bills and receipts arrive, with local OCR and AI analysis before anything is booked.
Document processing queue, sample data
Sample dataEvery document state, with held jobs waiting for a person.
Write a check, sample data
Sample dataSay who, how much and why once, and the check is filled in for approval.
Office overview on a phone, sample data
Sample dataThe office overview on a phone.

03 / Architecture

How the pieces connect.

Next.js 16ConvexClerkClaudePython OCR workerPlaywright
ScannerIntake workerOCR, QRLedgerConvexClassifierproposesReview queueperson approvesOwner Q&Apaper, voice, tabletanswersScannerIntake workerOCR, QRLedgerConvexClassifierproposesReview queueperson approvesOwner Q&Apaper, voice, tablet
DeviceEdge / networkDataAIPeople / UI

04 / Result

650 → 24

held bookkeeping jobs after one review round. Held jobs went from 650 to 24. 388 owner answers were placed in one pass. The question queue loads in 1.9 s, down from 4.3 s.

To-do home

Opens on what needs a decision today.

Paper in, data out

Printed packets come back through the scanner and match by QR.

Checks with limits

Owner sets limits. Anything over needs a PIN.

Two readers

Scanned pages are read twice by different models and compared.

55 screens

All of them crawled by a signed-in test on every release.

Clear review queue

AI proposes, a person approves. Nothing posts on a guess.

Have a problem like this one?

Book a call

Next step

Tell me what isn't working.

A 30-minute call. You describe the problem, I tell you straight whether it's worth fixing and roughly what it costs.

Book a call