Sous is a generative AI cooking assistant that builds meals around the ingredients you already have, does the nutrition math you were never taught to do, and gets out of your way the second you don't need it.
At a glance
Role: Designer (team of 4 on the original build; solo on the 2025–26 revamp)
Timeline: 2024 (as Perfect Plate) → 2025–26 (rebuilt as Sous)
Methodology: Design Thinking (double diamond)
Tools: Figma I owned the nutrition tracker end to end - flow, mobile screens, and the tablet redesign - and led the design-system rollout across the whole app. I co-led usability testing and shaped the interview protocol that pointed us in the right direction early.
Why Sous
You already know this feeling. You open the fridge, it is full**, and somehow there is still nothing to eat.** Not because the food isn't there, but because turning "stuff on shelves" into "a meal" is its own small act of creativity, and some nights you just do not have it in you.
Every recipe app I tried solved the opposite problem. They assumed you already knew what you wanted and just needed the instructions. Nobody was solving the actual mess of weeknight dinner: the ingredients you happen to have, the nutrients you happen to be short on, the energy you happen to have left.
So we built the thing we wanted to exist. Sous starts from your ingredients, not the algorithm's greatest hits. It tracks nutrition based on what you actually ate, not what you meant to. And it hands the decisions back to you - including hands-free voice, because your hands are covered in garlic and that is the entire point.
The name is the whole philosophy in one word. A sous chef preps, suggests, and stays out of the way. They do not shove the head chef aside and take over the pass. Neither does ours.
Approach: assistive, not authoritative
Here is the instinct we had to fight the whole time. Every AI-first product wants to do more. Suggest more, log more, generate more, decide more. But our own research kept pushing back. When we talked to the people we most wanted to reach, a clear group emerged - the ones who genuinely love to cook - and they told us the same thing in different words: do not take the wheel. Do not flatten the parts I enjoy. Leave me the discovery, the improvising, the deciding-for-myself. They did not want a robot chef. They wanted a really good assistant.
That finding is the whole reason we committed to Sous being an actual sous - and it forced a sharper question than it first sounds: what is the real difference between an app that takes over cooking and one that assists it? On the surface they look identical. Both suggest recipes, both track nutrition, both walk you through steps.
The difference is who holds the pen. A takeover app decides for you - it picks the meal, runs the plan, optimizes you toward some target, and your job is to follow along. A sous sets you up to decide - it preps the tedious parts (what is in your fridge, what you are short on this week, a few options worth a look) and then steps back, so the choice, and the fun of making it, stay yours. The test we held every feature to was almost embarrassingly simple: when you close the app, do you feel like you cooked, or like you were cooked for? The moment it was the second one, we knew we had gone too far.
So we ran the full double diamond - empathize, define, ideate, prototype, test - but the questions the framework kept surfacing were about restraint, not horsepower. What should the AI do without being asked? What should it never do without being asked? Where exactly does "helpful" tip over into "took over"?
That gave us the line everything else hung from: the AI's job is to give the cook more options, not fewer decisions. Every hard call downstream came back to that one sentence.
What I designed
Nutrition tracker. This one was mine, start to finish, and it is the feature I am fondest of. Most nutrition apps pick one of two ways to fail: they infantilize you (traffic lights, gamified guilt) or they bury you (raw macros, zero context). Our cook wanted neither. They wanted awareness without judgment.
So I designed nutrition in relationship to the week's cooking instead of against a rigid daily target. What did you eat today, this week, this month? Where are the gaps? What is already in your fridge that would fill them? The tracker feeds straight back into meal suggestions, turning "what should I cook" into "what should I cook given what my body's actually had this week."
And there was a quieter problem underneath the visible one. Most people know they are supposed to eat well. Far fewer know how to turn that into numbers - what a day of protein looks like on a plate, whether they are low on iron, how a single meal adds up. So Sous does the arithmetic for them. It calculates the nutritional load of what you cooked and ate, then names the gaps in plain language, so "eat healthier" becomes "you're low on X, and the carrots already in your fridge would fix it." You bring the ingredients and the intent. The tool does the math.
For the tablet redesign I grouped nutrients semantically - macros, vitamins, minerals - with a horizontal stacked bar reserved for the macro view. That quietly solved the "too many nutrients in a lopsided grid" problem before it could become one.
Design system rollout. Sprint three brought the chaos every team pretends won't happen: screens by different people were drifting apart, visually. I led the consolidation - locked the palette, the type scale, the spacing, the component library, then rolled it across every screen: auth, chat assistant, suggestions, tracker. Muted sage as the primary, warm cream as the surface, terracotta for state. Fraunces Bold Italic for the editorial moments, Inter for everything else. It is not glamorous work. It is the exact difference between a coherent product and a folder full of designers' opinions.
Usability testing. I co-led two rounds of think-aloud testing across low-fi and hi-fi. The discipline was simple and unforgiving: pick tasks that stress your assumptions, watch without steering (think-aloud study), and pay closest attention to the moments where a user invents their own logic instead of following yours. Those moments are where the real design notes live.
The revamp: from Perfect Plate to Sous
Here is the thing nobody tells you when you design a cooking app: you probably designed it for the wrong screen.
We built the first version, Perfect Plate, in 2024, and we built it the way everyone builds everything - for mobile. But picture how you actually cook. Your phone is not in your hand; your hands are busy. You glance at a screen a few feet away, between chopping and stirring, like a recipe card propped against the backsplash. That is a tablet. So in 2026 I tore it down and rebuilt the whole experience for the screen a sous chef actually lives on.
None of that would have worked without the groundwork the first version laid. Perfect Plate was the full Design Thinking method with the receipts - personas and empathy maps, interviews pairing Likert scales with open-ended questions, thematic analysis and affinity diagrams, How-Might-We statements, Crazy 8's, think-aloud testing from low-fi to hi-fi. That is the work that taught the product what it was for.
Then I re-dressed it. The old palette was already green, but it was a little flat and a little scattered, doing the color no favors. I pulled it toward a softer, more cohesive green and tightened the entire system around it, so the interface finally feels the way the product is supposed to feel: calm, health-conscious, alive. That redesign is the moment Sous stopped looking like a class project and started looking like something you would download.
Reflections
Detachment from your own work is a skill, not a mood. Watching a user fumble through the tracker in testing, I felt the pull to jump in and explain. That pull is the tell: if a design needs my voiceover, it is not done. Sitting on my hands in silence is one of the hardest muscles I have had to build.
__
AI does not remove design decisions. It multiplies them. Every capability the model adds is a question I now have to answer: does the user even want this? Automatic or on request? What does the screen look like when the AI is flat wrong? Sous taught me that designing for generative AI is less about the model and almost entirely about the guardrails you build around it.
__
Systems thinking is design thinking. That sprint-three design system shaped the final product more than any single screen I drew. Locking the palette, the type, and the components was the decision that let everything after it feel like one product instead of a truce between four opinions.