itssujeeth Principal Product Architect · 2019 Demonstration
Deep-Learning Product Profiling
A computer-vision demonstration that reads product attributes from packaging imagery, built to show how manual catalog tagging could be automated.
Onboarding new products into a retail catalog required slow, manual attribute tagging from packaging and product photos. The question was whether a vision pipeline could do that work reliably enough to be worth productizing.
Designed an image-processing pipeline using pixel-intensity detection, segmentation, and classification to read product information from packaging, and delivered a working demonstration with a product roadmap.
Delivered as a demonstration and roadmap. It did not progress to deployment, and no commercial outcome is claimed for it.
The interesting constraint here was working at the intersection of two unreliable inputs: product imagery is often low-quality, inconsistently shot, and poorly lit, while the attribute taxonomy it needs to map to is business-defined and changes over time. A model that performs well in controlled conditions often breaks in production when the photography varies. Building a pipeline that stayed reliable across that variation was the real engineering problem. Achieving high accuracy on a clean test set was the easier, less relevant goal.
This one stayed a demonstration. I include it because the line between a prototype result and a production claim matters to me, and a portfolio that only lists wins is not an honest portfolio. The pipeline worked on the demo set and the roadmap was credible, but the business did not fund the deployment, and I do not claim an outcome it never produced.