AI PRODUCT DESIGN, USER RESEARCH

SOUS-PILOT

A kitchen assistant that schedules the cook rather than the recipe — turning a
multi-part meal into one timeline that adapts to how fast you actually work.

PROBLEM:

Recipe apps solved information and left orchestration alone. A recipe is written for a stranger, in an order set by someone who has already made the dish, at a pace nobody measured. That holds up until a meal has more than one component — and then the cook is doing scheduling work the format has no way to carry, with wet hands, against a phone that keeps locking. The easy diagnosis is that people don't know how to cook. Five interviews said otherwise: everyone knew what to do, and still ended up with a cold side, a protein that rested too long, and the worst minute of the evening arriving all at once.

APPROACH:

I interviewed five cooks, from a former line cook to people who cook a few nights a week, and asked each about the last multi-part meal they actually made rather than about cooking in general. Four failures recurred: components finishing minutes apart, prep estimates that run about double, screens that lock mid-step and cost a hand-wash to reach, and a spike of stress at the very end when plating, cleanup and the table all land together. Sous-Pilot treats the meal as a project plan instead of a stack of recipes — it merges the steps into one critical path, learns the cook's real pace instead of assuming the recipe's, and re-derives the remaining timeline silently when they fall behind rather than flagging them as late. The setup screen offers the pacing as a choice (relaxed, efficient, or everything-hot-at-once) because the right answer changes with how much attention the cook has that night. Interaction is voice-first and glanceable, since the research made clear the screen is itself a failure point. And every timing prompt carries its reason — starting the oven now, it takes twelve minutes to reach 400° — so a wrong instruction is visibly wrong before it costs dinner rather than after.

OUTCOME:

A defined concept, an interaction model, and the two screens that carry the load: the setup where pacing is chosen, and the heads-up display that runs the cook. What the project actually settled is narrower and more useful than the concept itself — that the unit of help in a kitchen is the schedule, not the step, and that an assistant's credibility rests on explaining its timing and absorbing the cook's pace without comment. What it didn't settle is the finding that cut against it: asked what an assistant should do, four of five wanted help deciding what to cook, not when to start. I kept the timing focus because timing is what they described going wrong. That gap is the first thing I'd test.

SCHOOL:    Stanford Online

COURSE:    UX/UI for AI Products

ROLE:        Research, Concept, Interaction Design