Debt Hacker · Fintech app, not yet launched · 2024 to now

Proving the money math before anyone acts on it

For people who owe more on their credit cards than they can pay down fast

With help from Patrick Hrabos

The result

Caught by our own audit

An AI-written test report called the math accurate to the penny. Our audit against an independent calculation found a savings figure 26 times too high, in a step the AI-run tests had listed as never tested.

The short version

The problem

Advice people act on

Someone told to pay $500 more on one card will do it, so every figure in the plan has to be right, starting from statements that every bank lays out its own way.

What we built

AI reads, code calculates

Our data pipeline turns a person’s statements into one plan for the month: AI reads the paperwork, the person checks what it found, and the math runs in code that can be checked.

What a simulation suggested

Simulated: about 4 in 10 → 7 in 10

In a provisional simulation of 3,000 households that we had AI agents build, those stuck in debt got out within ten years far more often with the approach in full, mostly thanks to parts still to be built.

From a pile of statements to one plan

We built the data pipeline and AI that turn a person’s card and bank statements into Debt Hacker’s plan for the month, the same kind of work that has to happen before any company’s AI can act on its numbers. Jeremy Brooks, who founded Signal Scout, invented the app.

We built it the way data teams build ELT pipelines, in layers. Statements arrive as PDFs or CSV exports in many different layouts, and AI reads the PDFs and any CSV layout our code doesn’t already recognize. From there the pipeline cleans the raw records into accounts and transactions and finds the recurring bills and income in them.

Then it works out each card’s interest, minimum payment and the cash left over each month. It compares several ways to pay the debt down and combines them into one plan, in more than forty steps that each run once their inputs are ready.

The first version ran as dbt models in a data warehouse. We moved the same layers into the app’s own backend, so each person’s plan is rebuilt on demand and their data sits under their own sign-in, apart from everyone else’s.

How statements become a plan

  1. Card and bank statements, as PDFs or CSV exports
  2. AI reads them, and the person checks what it found
  3. Raw records cleaned into accounts and transactions
  4. Bills, income, interest and the cash left each month
  5. Ways to pay down the debt, compared
  6. One plan for this month

Where AI works, and where code does

AI reads the statements, and the person sees everything it found and corrects it before the plan is calculated. Once the plan is ready, AI explains it in plain English, working from the stored results with no way to change a figure. Every calculation in the plan runs in plain code that can be checked, and if more money goes out each month than comes in, the app refuses to make a plan and says so.

Checking the math against a second calculation

AI agents wrote most of Debt Hacker’s code, and an AI-written test report said every step of the pipeline had been tested and its math was accurate to the penny. We had AI agents audit the pipeline against a second calculation, built a different way from the pipeline’s own, and the audit found a $12,000 test balance being told it could save approximately $181,835 in interest. The true figure was $7,089, and the step behind it was one the AI-run tests had themselves listed as never tested.

The audit traced its errors to copies of the payoff math scattered across the pipeline’s steps, so we replaced them with one shared calculation. We ran 1,684 checks on nine household profiles built from published figures on card debt, and the new calculation never disagreed with the independent one or broke a rule any right answer must follow, such as savings never exceeding the most interest a balance could ever cost. Then we replayed every error the audit had found, this one included, and the new checks caught each one.

One $12,000 test balance

$181,835

what the pipeline said it could save before the audit

$7,089

the true figure, from the independent calculation

Testing the approach on 3,000 simulated households

To see whether Debt Hacker’s approach would work, and for whom, we had AI agents research published US data on debt, income and how people really pay their cards, then build a simulation of 3,000 households over ten years. For households stuck in debt, those owing $3,000 or more they couldn’t clear within a year, the approach in full raised the share who got out of debt from about four in ten to about seven in ten, and most of that gain came from parts still to be built. These numbers stay provisional until real people use Debt Hacker, and Jeremy wrote up how the simulation was built.

We can build pipelines like these for your AI, and check their numbers against your own systems before your AI relies on them. See how we work with companies whose AI runs on their data or with software companies.

Next story

Two stays in the Texas Hill Country, run at a profit

Lost Horizon · Founder · 2021 to now

Read the Lost Horizon story