Spearheaded the design of AI-powered user journeys, optimizing store operations, reducing out-of-stock rates by 10% in test stores, and saving an estimated $15M in revenue. Pioneered a strategic human-in-the-loop framework, optimizing data labeling tasks for a 91.7% CSAT.
With 100 million IP cameras shipped annually, computer vision technology holds tremendous potential for retailers. Recognizing the need for a fresh perspective in the retail industry, Store Vision AI Services aimed to enhance customer interactions, streamline operations, and stay ahead of the competition. Collaborating with leading retailers, we redefined the in-store experience to help store crews to improve their store operation.
The project began without a well-defined vision for how the store should operate. To address this, I partnered with our researcher, Obinna, to directly observe store managers and crews.
Then I translated the insights into journey mapping and storyboards to fully explore the problem space:
I designed both in-store and data labeling experiences, creating a unified ecosystem.
Store crews leverage AI to prevent out-of-stock events and optimize store operations.
Human labeling enhances the accuracy and performance of AI models.
I drove the creation of a user-centered AI framework, improving efficiency and aligning stakeholders around real user needs.
Store managers want new technology to improve their work. We need to be honest about what AI can do – it's not magic, but a powerful tool that works best alongside human knowledge. That's why I designed the homepage to show this partnership. Managers will see that AI is strong, but needs their expertise to get even better.
Priority tabs streamline store operations for managers, focusing on their top concern: preventing out-of-stock situations. To foster trust, I balance powerful AI with transparency. Camera views are emphasized, allowing managers to verify results easily. If needed, they can provide feedback, helping the AI continually learn and tackle even challenging scenarios like fridges or enclosed shelves.
To balance data labeling speed and accuracy, I employed a research-backed approach using 3 rounds of user testings. One round worth to call out is to explore how AI assistance could boost both the speed and accuracy of data labeling.
My ideas worked! Just giving explanations didn't help much, but using predicted labels made things way better. Adding a confidence score made it even faster and more accurate.
The final UI emphasizes simplicity: one item at a time, alongside predicted labels, confidence scores, and an insights panel for maximum efficiency.
This is the Experience Fluctuation Model. It shows how a task's difficulty should match someone's skill for the best experience. To make labeling tasks more engaging, we can make them slightly harder as the user gets better. This helps them stay focused and interested.
The AI tracks individual performance, offering skill-building challenges at the right time for continuous growth.
The AI recognizes patterns in performance, suggesting breaks or resources when needed to optimize the workflow.
Data labeling work can be repetitive. To combat fatigue and boost wellbeing, I focused on:
Provide engaging insights throughout the process, connecting labelers to the greater impact of their work.
Celebrating success means recognizing those who go above and beyond, showcasing their wins, and fostering a culture of achievement.
Helping labelers avoid burnout by proactively suggesting rest periods after long stints of focused work.
