Problem
Brady Corp's lab-materials teams manage thousands of container SKUs — beakers, bottles, jars, carboys — and labeling and inventory workflows depend on knowing each container's volume. That metadata is often missing, inconsistent, or only recoverable by manually measuring physical samples, which stalls product-data work.
Approach
- Segmentation network: isolate the container from the background in a photograph so measurements aren't polluted by whatever the item was photographed on.
- Regression network: map the segmented container image (with a reference object for scale) to a volume estimate, learning the shape → capacity relationship across container families instead of hand-coded geometry.
- Desktop packaging: a polished application rather than a notebook — camera capture, inference, and confidence reporting in a workflow lab staff could use immediately.
Outcome
The system predicts container volumes from a single photo with roughly 90% accuracy and took 1st place at the MSOE Innovation Lab Competition, Brady's internal showcase where student teams present to engineers and leadership. The project sharpened the exact skills my hiring managers later cared about: taking a fuzzy business ask ("get us volumes"), building a computer-vision pipeline end to end, and demoing it to a non-technical audience.
Related
My later research on 3D manufacturing perception — cross-category zero-shot anomaly detection on point clouds — continues as OptiQual3D on GitHub.