AI in Engineering, Pricing, and Data
We did not rewrite anything. Checkout still dropped from 9 seconds to 5.
A regional grocery chain had a loyalty app and a small ops team. When lines backed up at peak hours, everyone blamed the database.
So we added a cache, tuned indexes, and pushed more servers. Latency improved a bit, but the bill and the on-call noise both went up.
Then we traced a single “price check” tap. It triggered 14 backend calls, most of them fetching data the screen never showed.
The real issue was not “make it faster.” It was “why are we calculating promos, recommendations, and inventory confidence on every tap when the user just needs a price and aisle.”
We removed work: stopped recomputing the same basket totals, killed duplicate lookups, and delayed anything not needed until after the scan.
This month, those extra calls are not just slower - they are visible line items in cloud and monitoring costs.
The biggest win came from deleting three endpoints and one background job, not adding a clever layer.
It makes me wonder how much of our performance problem is actually unnecessary work wearing a technical costume.