Diary of an AI Engineer

My AI judge caught my AI coach getting the rules wrong ten times in one board game night. The next night a coach caught the judge. My agents shipped 21 versions of the game app while we played.

An AI agent reviewed my 34 AI agents and said the bottleneck was me. They had asked me 110 questions in one week. I answered yes 62 times. So I changed what reaches my phone.

A two minute note from August told me exactly what to measure when my roof bars arrived in October. They came without a channel on top, as warned. AI read ten shops in minutes, and the fix cost about twenty lei.

When I added one listing, a year of payouts in my rentals tool landed on the wrong properties. Fifty two of them. Elsewhere, a scheduled job had failed twenty four thousand times since January, and nothing ever told me.

Reading eight rows by hand stopped a backfill that would have moved about a million pounds the wrong way. Seventy seven of the hundred flagged records were not errors, just agreements that ended early. My read only credentials would have blocked it too.

One copied alias line stopped a search index from updating for about four hours. My check passed because it used a path that skips validation. Then the repair script reported one hundred records dispatched, three times, and processed none.

My AI research said twice that a wheel for my off roader did not exist; it was a factory spare. A tape measure settled the brake question in four minutes after it flipped twice in one evening. A winch noted at 4,700 lei cost 7,000.

For two weeks our main branch stayed green while nearly every pull request failed its end to end check. The suite never ran on main. One failure dismissed as flaky was real: users had lost their support chat after login.

A rule my agents follow ninety percent of the time still breaks four times a day. So rules are moving into code: one guard refuses edits to about thirty config files. Another lets work through when it cannot check.

My agent reported the test suite clean when the suite had never even started. A refused database connection produces no failing tests. Adding weaker verdicts, like partially verified, made the exaggeration largely disappear.

Deleting one hard filter improved my search results more than anything I did to the model. Ask for a platform engineer and it silently dropped the site reliability people. One provider's error messages turned out to be its best documentation.

A desktop app crashing on startup took four version numbers in one afternoon to fix properly. The first fix only stopped the symptom. The worst bug that month was an updater that reported success and left users on the old version.

On the twelfth of May, an author that is not a person landed a commit in our repository. By the following week it had four merged commits. Reviewing them changed how I review code written by people too.

Turning my agents' prompts into plain written procedures deleted about a third of the lines. Each one is now a document a new contractor could follow on Monday. When a run goes wrong, the fix goes into the document.

I gave the agent that works my backlog the ability to change its own source code. At once it replied to its own messages and wanted to restart after every change. It still cannot merge: it opens a pull request like anyone else.

The chat part of my bug fixing agent took an afternoon, and making it usable took the week. One run can take two hours. Long model sessions made the orchestrator decide the worker was dead, and the fix was four lines.

Static analysis on a codebase from 2016 produced about six thousand findings, which is the same as zero. Sorting them turned up 231 references to classes that no longer exist, each one a line that throws when it runs. Agents did the batched fixes.

In eight days I built a tool that reads a contract and tells you what is missing. About ninety commits, two languages from day one. There was no discovery phase: every question it asks is one I once paid for not asking.

I made 120 commits in a month on a tool with exactly one user: me. It runs my short term rentals. I looked at its revenue breakdown maybe four times, but I ask every day if the flat is ready for tomorrow.

Forty five scripts read a shop's stock level, and one of them quietly wrote it for months. Nothing errored, because both systems did what their code said. Counting the writers of every value has found a problem on all four systems I tried.

Five stages of a shop project were marked complete before lunch, and the basket still did not work. It was the busiest day in the project's history, 43 commits. Making the separately written pieces agree took the rest of the day.

I built five working products in four months, and only one has users who do not know me. It was the appointments platform. It won on boring work: Romanian copy, timezones, emails that do not look like spam, and telling people it exists.

My spec for a small shop asked for microservices and 80 percent coverage, and it shipped with zero tests. Six days later I started a shop for a real business that ran for ten months. Every spec I write now opens with two lines.

One file of project instructions changed my work more than anything else since March. It holds the commands, the conventions and the traps. Six weeks later, the first person it saved was me.

I built a sonar app in seven days and tested almost all of it at a desk, not on water. Two hours of recording the device's real stream paid for the week. A depth of zero, it turned out, means no bottom was found.

AI wrote most of my first video pipeline: 93 commits in five days, 49 of them on the Friday. Twice that week, code I had accepted worked by accident. A fifteen second redraw test now decides what goes in.