My role: Principal Product Designer, the only designer on a team of 1 PM and 3 engineers. I owned the proposal and approval flow, the AI guardrails, and the interface copy.
Launched October 2026.
Operators approve a ready-made fix instead of writing a ticket, so broken lockers reach the right team sooner.
| Before | After | |
|---|---|---|
| Time from issue to ticket | Before 3 days | After 3 to 6 hours |
| Usability score (UMUX, out of 100) | Before 65.1 | 71.6 |
Time to ticket is an early signal we're still monitoring: offline lockers used to surface after about 3 days through driver activity, customer activity, and support issues. The same events now raise an alert as they arrive, usually within 3 to 6 hours. UMUX compares the end of Q2 with the end of Q3 2026, before and after launch.
Depot and regional operators found broken lockers through driver reports or by checking the portal by hand. Writing every ticket by hand delayed repairs.
I tested my code prototypes with depot operators.
Propose
Vera explains what happened, the evidence behind it, and what it affects, so operators can judge the reasoning before they act.
A decision I reversed
My first draft rated every alert, for example "Confirmed, 93%". A score asks operators to trust the system's certainty. Evidence lets them check it, so I replaced scores with what happened, the evidence, and what it affects.
First draft
Final
Approve
No ticket goes out without an operator's approval. Once approved, Vera raises the ticket in nxt and relays the provider's replies.
Every plan step says what it does and why. Nothing runs without the operator's approval, and steps that can't be undone ask for a second confirmation.
Keyboard first
Verify
An activity log shows each ticket's progress as the provider replies. Operators get a notice on any page when the repair is done and the locker is back in service.
Repair news reaches operators wherever they work, at the foot of the navigation, instead of waiting inside the tracker.
When it's unsure
When the evidence is thin, the proposal says what the data can't rule out and gives a safe default.
When it's wrong
A rejected plan never runs. The operator gives a reason and can still write and raise the ticket by hand.
Process: Discovery interviews with 9 depot and regional operators, 2026. Card sorting ranked which alerts to build first, and 1 validation test checked the prototype.
I prototyped every interaction in code and tested it with operators. Figma came last, for refinement before engineering shipped it.
I spent several rounds polishing how confidence looked before testing whether operators needed it at all. They didn't. Next time I'd test the words and the decision first, and the visuals after.