Engineering case study

Kaevo

A household operating system. Built solo. Running in production.

Kaevo is a personal financial and household management platform: budgeting, cash flow, net worth, retirement Monte Carlo simulation, property tracking, travel planning, a family calendar, a document vault, and an AI assistant that works across all of it.

It started as a single HTML file. It is now more than 100 JavaScript modules, with payments, bank data aggregation, encrypted credential handling, a real pre-production environment and a gated release process.

It runs in production for one household. Mine. It is not open for new users.

Why this page exists

Kaevo launched publicly in August 2026 and did not find a market. I stopped selling it and kept running it, because the product was never the only point.

I wanted to know what it actually takes to build and operate a real system with AI coding agents as the engineering capacity, and what has to change in how you work before that is safe rather than just fast.

Four months in, the answer was clearer than anything I had concluded about it from the outside.

What I found

  • Throughput stops being the constraint. Within two weeks, writing code was no longer what stood between an idea and a shipped feature. Judgment was.

  • The ratio between production cost and verification cost inverts. Producing a change became nearly free. Confirming it was correct cost exactly what it always had.

  • Architectural entropy accelerates. I finished with 96 modules and a running list of functions living in the wrong files. Nothing was broken. Everything was slower to change than it should have been.

  • The worst bug I shipped was a save function that swallowed its own errors. Everything reported success. Nothing persisted. It reviewed beautifully.

The rules this runs under

  1. The agent has no production database access. Not read, not write. Every production data operation is executed by a human.

  2. Production promotions are human-executed. The agent proposes. A person ships.

  3. Every fix ships with a negative and a positive control. A test that fails before the change and passes after, plus proof that what should not have changed did not. A fix without both is a claim, not a fix.

  4. The process fixing a defect is never the process with authority to loosen the check that found it. If the fix and the test proving the fix come out of the same untrusted pass, you have built a check that can rubber-stamp itself.

I wrote the full operating model up here, including what I would do with it running an engineering organization:

Read the essay An Operating Model for AI-Native Engineering

What's in it

The Kaevo dashboard for a demo household: a greeting with the local weather, four cards for balance, spending money, net worth and investments, the House Score of 91 broken down by area, and this month’s spending by category.
The daily briefing. Balances, the House Score and where the month’s money went.
The Cash Flow forecast: a running balance chart for the month and a dated list of upcoming bills and income with the balance after each.
Cash flow forecast. The month’s running balance, and every bill and paycheck still ahead.
The retirement plan: the chance the money lasts to age 90, each person’s retirement date, spending and Social Security choices, and a projection of savings by year with a likely range.
Retirement plan. A Monte Carlo projection across 1,000 simulated markets, with the odds the money lasts.
The family calendar in month view: each day’s events, bills, meals and chores, coloured by household member.
Family calendar. Events, bills, meals and chores for each person in one month view.

Notes from building it

Michael Shollenberger

Who built it

Michael Shollenberger. Twenty-five years in engineering leadership, most recently VP of Software Engineering at Regent Education, leading a 28-person organization on a platform processing more than $2.5 billion in annual financial aid under Title IV compliance. Before that, enterprise technology strategy at Goodwill Industries International. I started out founding a software company and selling it.

Kaevo is what I built when I wanted to find out whether the AI-native engineering argument held up when I had to do all of it myself.