AI Development in Practice

More than twenty thousand lines of code, and the AI told me it was all done. Then I sat down to review the Kafka setup, and it admitted the most critical part was never built.

I check which tests actually ran, whether they used the application's real configuration, and whether they would catch the original defect.

Before I accept an AI-assisted query optimization, I check the real column types, returned records, and what changed in the execution plan.

I asked AI to change the code while leaving the tests alone. Swallowing an error and changing an acknowledgement setting produced different results.

The transcript was ready, but two waiting functions still held the recording. I traced those references and checked what could be released without breaking retries.

An AI-assisted review found retries blocked by deduplication, offsets advancing past failed work, and alerts lost after state updates. Before fixing the code, someone has to decide which outcomes are acceptable.

I wanted to use AI to organize component replacement reviews. Before writing queries, I had to understand why specifications, conditions, sources, and missing information needed to be modeled separately.