STEALTH
The fashion industry was drowning in waste—30% of inventory unsold, floating through discount bins or landfills. Mid-sized brands, clinging to quality, were being crushed by giants like Shein and Temu. Demand planners had to predict and place orders on micro-trends months in advance. One start-up emerged with a question: Could an AI-driven co-pilot save them?
I joined as Design Lead on a 4-month mission to find out.
My role
Worked with founder to scope the product sprints and design to optimize inventory decisions
Designed dashboards (low-fi → hi-fi) and shaped Ansel’s branding, all while collaborating with engineers to rebuild the backend during weekly sprints
Built a Google Sheets MVP as the front-end, bridging AI insights to brands’ workflows
Problem
The team had never prioritized the user experience of using the product for its 30+ year old, less digitally attuned users, and tasked me to embed complex AI into simple UI.
Impact
After shipping the redesign, we were able to increase engagement amongst existing customers
Existing customers:
50% of existing customers found this intuitively easy to use without an onboarding needed
50% of existing customers included this in their weekly SIOP meetings
Reduced decision-making time by 40% through actionable CTAs
Cut support queries by 25% (simplified, granular workflows)
Potential Customers:
2 potential customers said they would consider switching to our product for demand planning
Increased user trust in AI predictions (transparent explanations)
On-demand Planning
By talking to different fashion CEOs of mid-sized brands, we came to terms that in today’s hyper-season market, brands need to be adaptive in real time. With today’s micro-trends, demand changes every day, while brands need to make orders months in advance.
How can we demand planners with tools to make demand planning frictionless?
I first outlined the user journey across various jobs to be done throughout the season, then the information architecture unifies the team across various timelines.
The logic flow and 2x2 matrix prompt suggestive actions across different complexities and contexts.
All forecasting pointed to a function of risk
f(risk) = abnormal*movement
Working questions
How to best distill complex data at first glance?
How can we simplify information to what is absolutely essential?
How can we embed confidence levels in a way that is interactive and makes the user feel engaged?!
How do we seamlessly onboard end-users to reduce churn-rate after earning buy-in from leadership?
Considerations
Keeping data unbiased
Eliminate convoluted, unclearly structured data, or dark patterns early
Usable and scalable MVP
Standardized taxonomy - This builds trust and buy-in from demand planners by embedding familiar concepts that are fond to them.
Buildable over time
At a glance insights
↓ Number of clicks → ↓ interaction cost → ↓ mental load
Feedback Pathway
Provide feedback for the design as well as the accuracy of the forecasts and explanations themselves
With this new design goal in mind, I got to work exploring visualizations of the forecast data:
Data Visualisations
To show abnormality spring up over time, I experimented with various time series charts, this informed the product feed and card.
Some of hundreds of variations - had to pull out my dusty Econ book for this one!
The startup had live customers we met with at a weekly cadence. Continuous user-testing and feedback helped structure our product development, allowing us to show our new iterations and further refine the dashboard based on customer input.
By asking the customers to explain the chart to us and how they reached their conclusion, we were able to narrow down which types of visualizations directions were most intuitive to regularly scan and onboard onto.
Product Feed
Product watchlist
Help monitor at-risk products
Highlights extremes within size curve
Shows forecasted sell-through rate after 3 months
References Stock Tickers as a visual concept to identify movement and extremities
Widget Explorations
Adds contextualized insights based on JTBD
User-testing outcomes
Over the course of 3 weeks, we made 6 different iterations to the dashboard with the goal of embedding as much meaning in as few pixels as possible:
Forecast Value
Charts proved most easy to scan instead of lines or areas under the curve. So we used charts for the forecast value.
Below or Under target
Red and green could indicate whether the forecast performance is below or above target
Target Gap
The gradient distance from the target line could indicates by how much.
Final Designs
Before the redesign, there was lack of actionable or verifiable metrics. The dashboard was cluttered. The UI and explanations were cold, and overall inconsistent with the startup's branding and tone of voice. It was unable to hold up to the Start-up’s mission to make demand-planning lightweight.
Finally, the platform evolved from an initial draft into the final distilled version to help users identify and act on risk.
Development
Halfway through development, I adapted the final design into a responsive MVP using google sheets and sparklines. Back-end data would programmatically update into the client's sheet every day. This decision enabled the user to interact with the data, and onboard it into their sheet-heavy work-flow more seamlessly, while reducing development costs for such a lean team.
I recreated the platform by honoring type systems, using dynamic charts and spark-lines to code the data directly into the sheet as graphs.
Size-level Details
Break down size upon pressing the collapse to show the health by different sizes ensures this holistic detail is accessible and comparable on how the product is performing as a whole while eliminating necessary noise at the dashboard level.
Details on Hover
As users reported some confusion about the header copy when onboarding onto the new dashboard, we modified the note feature on google sheet to provide intuitive explanations upon hover.
In order to show it to users in real-time, we would develop the sheet one column at a time, from left to right. This enabled us to ensure the data was being interpreted smoothly, and iron out any miscommunications from data to final explanation.
This contractual role was a fantastic opportunity to engage in weekly sprints, ideate directly with product and engineering teams, and ship live changes based on real-time user feedback:
No such thing as too many variations
During weekly sprints, I committed to ideating a new component or data concept daily. By rigorously comparing options, we identified optimal solutions—a process highlighting why human designers remain critical in an AI-driven world—by articulating why certain hierarchies or components outperform others. This iterative approach also helped me detach from early ideas, ensuring only the strongest designs progressed to the final product.
Design against Waste
Everything is designed once, then used thousands of times
When exploring all the different ways to draw one forecast chart, the one thing that kept me going was the idea of James (name changed for privacy), a 40+ year old demand planning veteran, spending 8 hours every day staring at an excel sheet. Because of the start up's dashboard, he can save time by designing the data he needs to monitor in a way that is easier to find and interpret into actionable insights. So much of this world takes so long just to learn how to use, and it is the worst kind of waste.
AI ethics
The process of working with James and demand planners like him also touched on ethical designs for AI - where disclosures and notes needed to be in place to prevent overlying on AI-generated insights to inform multi-million dollar purchase decisions. I enjoyed narrowing down choices with James to reach the more intuitive dashboard we have today.
