Saws Weather Wardrobe App– Case Study
SAWS
Smart Weather
Assistant
An AI-powered mobile app that helps international students at US universities understand unfamiliar weather and dress with confidence every morning.
The gap between
data and meaning
Weather apps are built for people who already understand the climate they live in. A forecast of 28°F carries no lived meaning for a student who grew up in Lagos, Chennai, or Bogota. The number is accurate. It is not useful.
My role was to design a product that bridges this gap entirely: from the first onboarding screen through to the moment a student opens their front door each morning, confident in what they are wearing.
International student, University of Michigan, Winter 2025. This single research quote anchored every design decision that followed throughout the project.
Standard weather apps assume cultural fluency in local climate. They show numbers, icons, and precipitation percentages with zero frame of reference for users who have never experienced temperatures below freezing.
None of them personalise to where a user is from, only where they are now. That single omission is the core product opportunity SAWS was built to address.
What students
actually experience
I conducted interviews and observation sessions with international students across 4 US university campuses. Participants represented 11 countries of origin. Three dominant pain points emerged with consistent frequency across all sites.
Students from tropical climates had no experiential baseline for temperatures below 40°F. A number meant nothing. They needed comparison, not data.
Phrases like “dress in layers” carry different meanings across cultures. Advice that feels obvious to American locals is deeply ambiguous to newcomers.
Dressing incorrectly created a repeating loop: wrong outfit, discomfort, embarrassment, avoidance. Students began skipping class on cold or uncertain-weather days.
The problem was never the weather. It was the absence of translation between a number and lived human experience.
Version 1:
Testing the wrong assumptions
The first prototypes were built around a data-forward model: give users more information and trust them to interpret it. Three low-fidelity screens tested this with 8 participants across two campuses. More data produced more confusion, not less.
249am
2812pm
333pm
306pm
24
Warm top
Consider a scarf
Four rounds of
deliberate change
Each usability round surfaced a clear failure. Each failure produced a precise, tested correction. The product changed fundamentally four times before reaching a form that worked consistently across all tested personas.
Testing showed that phrases like “dress warmly” or “light jacket recommended” produced no meaningful change in behaviour. Students from tropical backgrounds had no reference for what those phrases meant in practice. Confidence scores stayed flat across all participants.
Switching to named garments with specific weight descriptors produced an immediate comprehension improvement. Students could match the recommendation directly to an item they owned.
Even with specific garment names, students did not trust the recommendation because they could not evaluate the logic behind it. People understand new information relative to what they already know. That insight drove the fix.
Adding a culturally-relative comparator increased trust significantly. I introduced the origin-country onboarding field specifically to power this comparison layer across all recommendations.
A student from Stockholm and a student from Accra need entirely different outputs for the same 28°F morning in Chicago. Early versions served identical recommendations regardless of origin. Round 3 caught this when two students with opposite backgrounds were tested together. One found the app overcautious. The other found it insufficiently urgent.
I designed a thermal baseline model: origin country defines a comfort band, and every recommendation is calibrated against it. Same weather. Entirely different output per user.
The profile system required origin country, wardrobe items, and sensitivity preferences. A 7-step onboarding had a 38% drop-off rate before completion. I collapsed the flow to 3 essential inputs, deferring secondary preferences to in-app settings accessed after first use.
The finished
product
Three screens represent the full user journey: onboarding, the personalised home dashboard, and the culturally-differentiated recommendation view. Every UI decision maps directly to a research-validated iteration.
Kwame (Ghana) and Lena (Berlin) see the same 28°F morning in Chicago but receive entirely different experiences. The same temperature reads as a weather emergency for one and a routine winter morning for the other. This is the core design achievement of SAWS: identical data input, personalised human meaning.
Measurable impact
across four campuses
After 12 weeks tested across 4 US university campuses, the results validated every core design decision made through iteration. The numbers reflect a genuine, documented change in how students began their mornings.
of tested students reported feeling significantly more confident in their daily outfit decisions compared to before using SAWS. Measured via end-of-day diary entries over 4 weeks.
US universities recorded a measurable drop in weather-related morning anxiety across follow-up surveys. Results held across tropical, subtropical, and equatorial student backgrounds.
Three things this
project taught me
-
1Specificity is a design tool
Vague language feels safe to write and design around. Named garments, concrete comparisons, and thermal baselines feel risky but they are what actually changes behaviour. The shift from “dress warmly” to “wear your heaviest down coat” was the single most impactful design change of the entire project.
-
2Personalisation must have structural depth
Surface personalisation such as adding a user’s name to a greeting does nothing. Structural personalisation, building the entire recommendation model around the user’s origin climate, changes the product entirely. The persona work in weeks 4 and 5 was what made the final differentiated screens possible.
-
3Drop-off data is as valuable as satisfaction data
The 38% onboarding drop-off in round 2 told me more about the product than any satisfaction score could. Watching where people leave is where the real design problems live. I now treat flow completion rates as a primary signal from the first week of any project.