Case studies

Key technical decision

Broke everything from preprocessing to inference into individual nodes, connected together

Preprocessing steps and model combinations needed differed by researcher, so a single fixed pipeline wasn't enough. Making each step an independent node that researchers could pick and connect as needed meant they could try new experiment configurations immediately, without a development team request.

Key technical decision

Did not auto-finalize reconciliation results; only exception cases required staff confirmation before closing

Since settlement involves real money, letting the system finalize automatically would delay catching errors. We split the flow so that transactions matching the rules were processed automatically, while anything outside the rules required a human to confirm and approve before closing.

Key technical decision

Had stores enter metrics themselves, with headquarters adding a final review step

Stores knew their own on-the-ground data best, but trusting entries as-is meant input errors would flow straight into the metrics. We added a headquarters review step after store entry, balancing data accuracy against the burden on store staff.

Key technical decision

Separated vehicle location collection from delay detection logic

Location collection needed to run continuously without interruption, while delay detection was a separate calculation factoring in route and schedule. Separating the two roles kept the location collection pipeline stable while letting delay detection logic be improved independently.

Key technical decision

Moved the whiteboard approach directly into a board-style screen, while designing the underlying data structure from scratch

Floor staff were already used to checking on process stages by moving items across a whiteboard. We kept the familiar interaction on screen while redesigning the underlying data to preserve process-transition history, so we could later trace which process stage delays repeated at.

Key technical decision

Kept the existing accounting system as-is, syncing inventory data through an API only

Rebuilding the accounting system would have meant re-validating tax handling as well, which was too risky. It was safer to build only the inventory management piece new and hand off just the figures accounting needed at closing, via API.