Automotive solutions company lands DMS with 6M images

Overview

A driver-monitoring system that doesn't see every face is a safety failure, not a feature gap.

A Global Automotive Solutions Company needed to train its in-cabin AI to detect eye blink, drowsiness, and attention states with consistent accuracy across skin tones, demographics, and lighting conditions.

Firstsource delivered 6 million+ images with 68-point facial-landmark annotation across skin tones ” capturing eye-blink status (open, closed, half-open) at a level the model could generalize from.

This was Intelligence that Operates: bias-aware perception data feeding a production driver-monitoring model from prototype through launch.

Challenges

• Driver-monitoring models fail on diversity gaps long before they fail on accuracy. A model trained on a narrow demographic distribution misses faces it should detect ” and the failure shows up in production as a safety issue, not a performance metric.
• 68-point facial-landmark annotation isn't a generalist task. Capturing landmark precision across skin tones, ages, and lighting conditions requires annotators trained on facial-feature precision ” not crowd-grade box-drawing.
• Prototype-to-launch acceleration depends on annotation velocity. A DMS model can't get out of the lab if the training-data backlog is the bottleneck. Annotation has to scale with model iteration, not behind it.

How We Made It Happen

We ran annotation at production scale with bias-aware capture and a single quality discipline.

• Facial-landmark annotation across skin tones at production volume. 68-point precision delivered across demographic variation, with eye-blink state tagging built into every record.
• Data Annotation + AI Safety (bias) in one workflow. Bias coverage wasn't a side audit ” it was baked into the labeling rubric itself.
• Annotation scaled with model iteration. Throughput rose 35% as the program matured, removing the data-side bottleneck on the client's algorithm-accuracy timeline.

Conclusion

A DMS model only ships when its training data sees every driver. Firstsource ran the annotation program at production scale with bias-aware capture ” turning automotive perception data into Intelligence that Operates.