Adding ADAS intelligence to older delivery vehicles

Thousands of delivery vehicles with zero collision warnings

As lead product designer for safety and innovation on Amazon's last-mile delivery team, I was tasked with designing a pre-collision notification system for a fleet of vehicles that predated modern ADAS features. Drivers operating in dense urban environments had no advance warning when a pedestrian, cyclist, or vehicle entered a collision trajectory.

The technical constraint shaped the entire design challenge: alerts had to render on screens that already existed in the vehicle — specifically the center console tablet that drivers used for routing and delivery tasks. I couldn't add hardware. I couldn't take over the screen. I had to find space within an interface drivers already depended on, placing notifications close to the road view so they'd register in a driver's natural sight line without disrupting their workflow.

This meant redesigning parts of the existing tablet experience to make room for safety information — before I could design the alerts themselves.

Making space for real-time alerts

Safety notifications needed to live as close to the road view as possible — that's where a driver's attention already is, and that's where it needs to stay in a critical moment. But the existing tablet UI was cluttered with persistent controls, expanded panels, and misaligned elements that left no room for an alert system. Before I could design the notifications themselves, I had to clean up the screen they'd live on.

From text to instinct: designing the alert language

With the screen cleaned up and a notification zone in place, the next challenge was the alert itself — how do you communicate a hazard to a driver in a fraction of a second? I explored three directions, testing each with drivers to find the one that required zero interpretation.

A simulator that replaced the need for physical vehicles

I built a web-based simulator that renders the full Rivian dashboard interior with real-time alert overlays. This enabled remote A/B testing of visual treatments, timing, and audio patterns — without requiring vehicles or controlled road tests.

A three-tier alert framework calibrated to cognitive load

The alert fatigue insight drove everything: alert intensity needed to match what drivers can actually process, not just how dangerous the situation is. I designed a framework with three severity levels — LOW (ambient awareness), MID (active attention), and HIGH (immediate action) — each with distinct visual weight, color treatment, and audio cues.

Alerts render on two surfaces simultaneously: the driver cluster for peripheral vision and the center console for detail. Eye-tracking data showed drivers check different screens at different speeds, which drove this split-surface approach.