Overseas Chief Quality Engineer Intern at Nissan North America, Smyrna. The team had 33,000 repair orders to sort and was doing about 2,200 a week by hand for monthly and early-detection meetings. I wrote a program to do most of that sorting.
My first plan was machine learning only, and there was not enough labeled data for it. So I learned how the group sorted by keyword and built that as a filter, which labeled 22,000 of the 33,000 orders. I trained TF-IDF classifiers on those labels and compared XGBoost, LightGBM, CatBoost, Random Forest, LinearSVC, logistic regression, naive Bayes, and SVM. The model sorted another 5,000, so 27,000 of 33,000 (82%) were sorted automatically, in under two minutes on a work laptop.
The engineers reviewed 1,400 of the results and found 1,300 correct. Most of the misses were orders that fit more than one group. It saved about 60 hours of manual sorting, and I added a GUI and a sheet and chart per car model so anyone could run it and change the groups without programming.
Wrote a 26-page handoff guide so later interns could install, run, and maintain the group's tools. Built a pipeline to decrypt, filter, and organize CAN bus signals, replacing an older department process.
The code and data belong to Nissan and are not public.