The Challenge: Gaining Real Visibility Into Model Behavior and Data Quality

Hexagon develops advanced computer vision models to support vital industrial applications. As their models grew more complex, so did the need to understand how those models behave — not just whether outputs were “right” or “wrong.”

Hexagon needed a way to:

• Uncover hidden patterns within their data
• Investigate failure cases quickly and efficiently
• Identify redundancies or gaps in datasets without manual review
• Trace how models “think” beyond black-box behavior
• Scale these analyses efficiently across large datasets

The Solution: Integrating Explainability and Data Insights Into the Development Process

Hexagon adopted Tensorleap to integrate explainability and dataset optimization directly into their computer vision workflows. Tensorleap integrates with the model to surface insights based on internal activations and learned representations.

• Deeper explainability: Tensorleap reveals how models perceive data by analyzing intermediate activations and surfacing patterns invisible to standard tools.
• Self-supervised data categorization: The platform groups data based on learned features, enabling automated detection of redundancies, gaps, and edge cases within datasets.
• Faster root-cause analysis: By clustering failure cases and similar behaviors, Hexagon’s team rapidly investigates why certain predictions fail and where improvements are needed.

Why Tensorleap Stood Out

• Limited scope: Other tools focused only on outputs without exposing internal model reasoning.
• Lack of flexibility: Many tools were tied to specific ecosystems.
• Maturity gaps: Available products lacked robustness or required heavy customization.

The Results: Improving Development Efficiency and Data Strategy

• 40% Dataset Size: Reduced redundant data and cut labeling and storage costs while maintaining performance.
• Faster Root-Cause Analysis: Isolated issues and failure cases faster, driving shorter development cycles.
• Better Data Decisions: Prioritized the right data for labeling and strengthened dataset quality.
• Streamlined Workflows: Integrated explainability directly into their development cycles.

Smarter, More Transparent Computer Vision Development

Tensorleap enabled Hexagon to move beyond reactive debugging and surface-level metrics, embedding explainability directly into their development process. By surfacing model behavior, data gaps, and labeling priorities earlier, they accelerated development, optimized datasets, and strengthened model reliability with confidence.