Rebuilding what 2,000+ trucks actually did, as it happened
Drivers were supposed to log every pickup, drop and delivery by hand, and often didn’t. I architected a real-time system that rebuilt each trip from the signals the trucks were already sending, and kept a person in the loop only where the evidence was thin.
- Role
- Architect and team lead
- Team
- Two engineers, daily code review
- Scale
- ~2,000–2,500 trucks on the road at once
The mess
Drivers skipped manual updates, dispatchers spent hours a day reconstructing what happened, and customers disputed pickup and delivery times.
The question
What actually happened on this trip, and which milestones can close without a dispatcher?
The call
Agree on the evidence before any agent reasons about it. We reconciled the signals into one trusted stream first, so the AI never had to reason about noise. And nothing was forced into the record: low-confidence events stayed open for a person. I wrote the core logic myself and reviewed the team’s work daily.
What changed
- ~90% of trip milestones completed automatically
- ~3–4 hours a day of dispatcher reconciliation removed
“A stop only becomes a milestone when the evidence agrees.”
Delivered within a prior employer’s environment. Details are intentionally anonymized. I’m glad to discuss the operating problem, design principles, and production tradeoffs.
Have a workflow that looks like this?
Let’s work out whether AI is actually the right lever.