Building thoughtful software with AI.Notes from the systems behind the work.
A good seller got an order of zero because its stock records had barely started, so a stockout looked like dead demand. I added a rule that reads the shape of the sales, backtested it, closed a second hole in the covered check, and caught my replay tool crying wolf.
Read moreA product that sold out the whole window showed zero sales, so the planner ordered zero and the shelf stayed empty. I added one bounded look-back for only those stuck rows, refused to invent demand when the product sat unsold, and labeled every row with where its number came from.
One vendor's slots matched nothing, so every slot got the same flat order suggestion. The stored size was a full label descriptor, the catalog left out the default size, and my existing default rule could never fire. Now the matcher pulls sizes out of the label, guesses the default only when nothing contradicts it, and prunes links a re-run no longer matches.
The home hub treated any remembered reading as proof a device was online, so dead plugs kept drawing power on screen. I moved presence onto the bridge verdict, chased it through every screen, and fixed a streaming box that had the opposite bug.
The discount tool counted package age from the producer's packed date, falling back to the last edit time. I had just bound its bulk write to a hash of the previewed plan; now age counts from the store's first receive event, unknown stays unknown and is never marked down, and the history read splits slow windows in half instead of failing.
One code path in my ordering system ignored the vendor's minimum-order setting and rounded every small shortage up to a full case. It now holds those lines for review with the real need still visible, still buys when a busy shelf is empty, and finally counts the back room.
A starter-case change for new products exposed that missing shelf counts and missing sales were being filled in as zeros, so readiness now checks how each number was obtained, the sales window keeps its first day, and only a narrow set of complete-window no-sale rows may count as a real zero.
Catalog rows are no longer bound by hand; a classifier does it, but only when exactly one candidate matches on every feature. Ties go to me, approval halts on any stray, and the first live run cleared a blocker without moving the blocked total.
A zero in my ordering plan can mean no demand or missing data, so I added a check that compares the plan against a raw summary of the store records, blocks the deploy only on a same-category contradiction, and warns on everything softer.
A store-lineup reader that treated any number as yes, a keyword guard that lasted under half an hour, a shape check that counts which kind of value sits where, a column fix a few hours later, and a word list that came back that evening.
The register feed and my spreadsheet catalog never shared IDs, so I put the links in their own table, link only on one exact brand-and-name match, and read sales through one route so nothing counts twice.
Letting my vendor credit report treat a digits-only promotion tag as a percentage was a mistake I pulled the same day; I now take the rate from structured data or work it out from the sale itself, with a headline that counts only claimable rows.
The new history panel in my hub had one tab with real data. Then I recorded on arrival and got lumpy lines, switched to a clock and got a recorder that can't tell a dead sensor from a calm one, and found a read side that can cut off the newest rows once the table grows.
Building a report to flag one product with two live package tags in the same store led to a day of fixing what it meant by the same item: an unmerged per-package feed, split rows added together, one size bucket for loose flower, selling-floor stock only, and product IDs instead of names. Unreachable stores still get skipped without a word.
The Unspecified card on my discount report filtered for the literal word instead of the rows with no supplier, which SQL can never match. A shared marker and one filter builder fixed the drill-down, with a few gaps I can still see in the diff.
Reading my coding-agent grader line by line, I found it stores friendly constants and stock turn values whenever the model judge fails, under the same label as real verdicts, so averages and the pass gate can't tell guesses from judgments; I want every score to carry its origin and every report to show coverage.
A store sheet had short, all-caps columns full of words, and my importer read them as store codes. The change checks a few cells under each candidate, hunts for the real header row, stops putting made-up prices into orders, and handles one cart crash. The new detection rules still have no test of their own.
Fresh branches were tagging sessions with the base branch's commit, so one error could match all of them. A merge-base check now drops that commit on the laptop, but the server's code-host lookup and backfill fill it straight back in for any pushed branch.
Every tool in a burst claimed the full cost of the one model call that followed it; I moved ownership to the call, then split its cost by tool output size, and the tests still don't check the math.
Switching long dashboard windows to weekly points went fine for counts and levels, but my handler averaged daily review rates instead of rebuilding them from counts, while the database path did it right. The two paths may also disagree about where a week starts, and short edge weeks and half-versus-half trend arrows add their own distortions.
A page I added scans stored agent conversations for leaked keys and personal data, and reading the diff showed five places where it drops rows, rules, matches, list entries or failed scans without a word in the response.
The dashboard guesses failed changes from pull request titles, and a substring match on the word fix, a loose time window and a mismatched denominator mean the number counts vocabulary as much as breakage.