Overview
Discoloration in ceramic production parts can indicate defects that affect part quality. Quantifying discoloration severity manually is time-consuming and inconsistent. This project delivered a production-deployed ML application that automates that assessment — giving the manufacturing group a reliable, repeatable tool for quality control.
The project was completed entirely during Internship 3, running from zero to first deployment in 3 months using Agile development practices.
What I Did
Agile Team Lead
Worked on a cross-functional team of 5 to design and ship the application. Served as scrum master for 2 sprints — running standups, managing the backlog, and facilitating retrospectives. This was the first Agile project within the R&D group at CoorsTek, and establishing that practice was as much a deliverable as the software itself.
Feature Engineering
Transformed 1,200 images into 1D feature vectors suitable for ML model training. The transformation pipeline extracted color and texture statistics that capture discoloration characteristics while keeping the representation compact enough for classical ML models.
Model Development and Improvement
Ran a series of experiments to diagnose why the initial model topped out around 80% specificity and F1. Identified data quality issues — labeling inconsistencies and class imbalance — and fixed them. Improved the model development procedure (better cross-validation, hyperparameter search, calibration). The result was a jump from ~80% to ~98% on both specificity and F1, establishing an MLOps standard procedure that the team can follow for future production deployments.