Building thoughtful software with AI.Notes from the systems behind the work.
An incremental sync that reset its coverage on every shape change, and starved its own connection pool, is rebuilt around batched transactions, a serial write queue, and a separate refresh cursor.
Read moreThe daily sales rate behind my order skipped the days the shelf was empty, so an order worth six to nine months of stock looked like two; now the rate comes from raw calendar days, the proof is pinned to the exact order version, and replay won't borrow another version's story.
A daily blind count, where the sheet hides the expected quantity, lost stock in its first version before anyone counted it, because a missing package quantity came through as zero and got skipped.
The per-store inbound field in my case planner was a hardcoded placeholder zero. I replaced it with a ledger of sent minus received units per order line, made old quantity-free partial receipts stop planning instead of guessing, and tightened the receiving endpoint so new receipts carry exact line quantities.
The case-ordering planner kept growing new fields under its version-one label until one tag covered three shapes; I froze the original, moved the new meaning to a successor, and made the planner stop and name any record caught in between.
A deleted-and-recreated cart line looked identical to my snapshot check, so reviewed checkout now trusts a database-owned version number and a one-time receipt instead of a comparison.
One continuity check guarded every feed in my store-data archive, so a delivery looked like corruption and a cleaned-up mismatch list looked like a collapse. Now each feed gets only the checks its data can survive, and a rejected snapshot gets another look in minutes instead of hours.
The task dashboard expanded a team into its members, lost every task the team itself owned, and fanned team assignments out to individuals. I kept teams as teams, filtered team views myself, and moved the end-of-feed check to the raw page so a page of other people's work can't end the scan.
A refunded pack whose credits were already spent had nowhere to go, so accounts got a separate debt counter and every incoming credit now pays that debt before it becomes spendable.
The save button always hit the create route, so editing a workflow left the original running and added a near-copy. Now the screen decides create versus edit up front, and the server appends a new version and moves a pointer instead.
The planner now pulls fresh stock before every order and refuses, with specifics, when a stocked product has no record, and a blank setting that had switched off the new request throttle got fixed too.
Widening sales capture to more stores left promotion with zero work. That zero was right: I had to make every join, done check and handoff in my ordering system compare the store, and then catch an overcount that showed up after the first repair.
A lookup in my ordering system built its list of already-handled products from any row with the right kind of ID, ignoring setup, brand, queue state and whether repairs were applied. I rewrote the query within the half hour, and the stub-based tests had to start checking the question instead of the answer.
The ordering system found zero sales rows for an approved product. Diagnosis showed one product was simply absent and another was filed under the wrong store connection, so I proposed an alias, counted its effect in memory, and only then wrote exactly one sales fact without saving the alias.
The first real approval in the ordering system was made as small as I could, with a couple dozen tables counted around it to prove it touched one thing. The counts held up. Then I found that several of the report's 'this didn't happen' lines were hard-coded, not measured, plus a few things counting can never see.
An atomic receipt store keeps each plan from my planning service exactly once, but the second guard ignored any plan that left its run field empty, in the database and in my own lookup.
When my vendor-credit report couldn't compute what a supplier owed on a promotion, it fell back to the shopper's full shelf discount. Now it derives the rate from the register's own amounts, applies it to cost, and writes zero with a needs-review label when it has no basis.
Sold-through packages kept landing in my planner's can't-find pile, so I taught it to check the report day's snapshot and the store's transaction history, matching only exact package IDs.
Widening a discount-and-return sync from one store to all of them meant the event ID had to carry the store too, while counters added the same day show what the merge absorbs and what the size cap cuts.
The sale tag in my point-of-sale system sits on a single package, not on a product, so I moved the on-sale report down to package rows and found the parts that still overcount.
The register quotes pre-packaged flower per gram, so my staff sale sheet printed jar prices that were short by the package weight. I now read the weight off the end of the product name, and the fallbacks around that parser still worry me.
A new screen turns a supplier price drop into a credit request and a register-update note, and the first version listed a product once for every store that stocked it.