Optimizing mobile screen taxonomy for always-evolving UI

The mission: Power faster UI innovation with AI-ready classification

A mobile app design company aimed to transform user interface design by training advanced layout-generation models to power faster, more dynamic UI innovation for its customers. To realize this vision, it needed support on labeling 117 complex design categories, allowing its internal team to focus on core AI development.

The challenge: High-volume categories, granular classification

Design taxonomies evolve constantly in mobile environments, and each screen element must be labeled accurately to train reliable models.

Key obstacles

• Complex annotation requiring precise labeling across 117 diverse UI component categories, which demands high design literacy and consistency
• The internal annotating team was at capacity constraints with the same number of resources divided between data labeling and model training
• Lack of a streamlined workflow that balances speed with consistent, high-quality annotations across a large and complex dataset
• Time-to-market pressures as data labeling challenges impacted the speed of prototyping and deploying new user interface experiences

The goal

Design a collaborative human and tooling environment.

The solution: Hybrid workflow with layered supervision

We built a robust annotation system optimized for high-volume, design-centric datasets that combined manpower and proprietary annotation tools to rapidly scale up labeling throughput.

Our approach

1. Leveraged our internal network of highly-trained and experienced labeling workforce
2. Implemented a scalable work distribution strategy tailored to the complexity and volume of 117 distinct categories
3. Utilized our proprietary data annotation platform to ensure consistency and clarity of deliverables
4. Integrated rapid and rigorous review control to amend errors before they accumulate
5. Leveraged our centralized platform to seamlessly coordinate annotators, tools, and reviews, ensuring high-quality data is delivered in client-ready formats

The results: Database expansion, ready for ingestion

• Scaled client's classified interface design database by 16 times
• Achieved 90% accuracy across 117 categories, exceeding the agreed service standards
• Freed the client's internal team to work on AI model refinement

By improving classification accuracy and streamlining their training data, we enabled the client to reallocate internal resources toward innovation and deliver greater value to their end users.