stmtscan: statements read by AI
Each PDF is split into pages and read with AI: bank, month, balances, deposits, NSF fees, negative days and every transaction, with payments to other lenders flagged. Under $0.01 for a 10-page statement.
named client case · internal AI tool · merchant cash advance
DocUp: the paperwork runs itself, and the costs are about $60k a month lower.
DocUp is an internal tool for the merchant cash advance industry. It takes in merchant leads, reads their bank statements, follows up by SMS and hands leads on to clients and their CRMs. The reading is done by stmtscan, an AI tool we built for it, and the rest by automations that took enough work off the team for DocUp to run with fewer people.
DocUp's owners run the business. I have worked on its technology since March 2021, with 651 of the repository's 979 commits, and we built stmtscan for it.
the original condition
About 4,500 statements a month, each one paid for, and a team doing the rest by hand.
A merchant's bank statements show what the business earns and who it already owes: deposits, balances, NSF fees, negative days and payments to other lenders. DocUp reads them for its leads. It was paying another provider $0.70 to read each 10-page statement, and its people were still checking, sorting and chasing much of the work around it.
what runs it
The automations behind DocUp.
Each one takes a step off the team. stmtscan is the one we built from scratch for it.
Labels, so people see only what needs them
Every scanned statement comes back labeled ready, other payments, inconsistent, failed or corrupt. A statement that duplicates one already on file for the same lead is removed.
Watermarks cleaned before reading
Statements with a watermark are detected and cleaned before they are read.
Lead imports that check themselves
Spreadsheets and ZIP files of leads import in the background, with a dry run first. Duplicates by phone, email or tax ID and blacklisted leads are caught, and the rejects come back as their own file.
Phones and emails checked on arrival
Every phone number is validated and sorted into cell, landline or VoIP, and every email is validated, before anyone calls or writes.
Follow-up and hand-off
SMS campaigns go out in batches on a weekday schedule, replies land in one inbox and STOP opts a lead out. Leads go to the client's CRM, DebtPayPro, Velocify or Centrex, or out as an export with a record of who received them.
by the numbers
DocUp's costs, and the repository's counts.
- to read a 10-page statement
- <$0.01
- Against the $0.70 DocUp paid another provider for the same work.
- statements a month
- ≈4,500
- The volume stmtscan was built to carry.
- a month off costs
- ≈$60k
- stmtscan together with DocUp's other automations, which took enough manual work off the team for it to run with fewer people.
- of 979 commits
- 651
- Mine, in DocUp's repository, from March 2021 to December 2025.
how it was built
From paying per statement to its own AI tool.
Build the platform
From March 2021: services for leads, imports, exports, SMS campaigns and three CRM connectors, with organizations, roles and permissions.
Bring the statements in
From October 2023: an integration for statement data from an outside provider, then blacklists, a reviewer queue and saved filters for the team working through it.
Replace the provider with stmtscan
A separate service on AWS that splits each PDF, reads it with AI, checks it and labels it. Merged into DocUp on 1 April 2025, with the cost of every statement logged.
Clean what the AI cannot read
April and May 2025: watermark detection and removal, and duplicate statements removed before they cost anything twice.
the stack
What it runs on, and why.
Node.js and Moleculer on AWS
Three services on ECS talking over NATS, MongoDB, and Lambda functions for the background work, in the Frankfurt region.
stmtscan on Lambda, S3 and SQS
PDFs split with Ghostscript and read with OpenAI, with a dead-letter queue that keeps any statement that fails.
Twilio, three CRMs, email
Twilio for SMS and its delivery reports, DebtPayPro, Velocify and Centrex for lead hand-off, Sendinblue for export links.
what this case does not claim
Named work, without invented performance theater.
- DocUp is internal: there is no public site to visit, and no client or merchant data on this page.
- The cost, volume and savings figures are DocUp's own. The $60k a month comes from all its automations together, not from stmtscan alone.
- The per-statement cost is for a 10-page statement. Longer statements cost more, which is why no monthly total is given.
- DocUp was started in March 2021 with other developers. I have been on it since that month and wrote most of it.
The paperwork runs itself now. What does yours cost you?
Tell me which documents your team reads, checks or retypes every day. In 15 minutes you know whether a tool can take it on and what the pilot would test.