Seita Fujiwara

Foreign object detection in alcoholic beverages: Fine-tuned and optimized a YOLO/MMDetection-based defect detection model with manually labeled data to significantly enhance accuracy. Constructed a mathematical tracking model integrated with the defect detector, raising classification accuracy from 83% to 93.2%. Refactored code and algorithms for better real-time performance while maintaining accuracy. Power distribution equipment failure cause prediction: Refactored existing code and added new features, improving the accuracy of a 14-class classifier from 77% to 82.8%. Processed and analyzed waveform samples using PCA, FFT, and clustering to identify root causes. Improved the accuracy of a 14-class classifier from 77% to 82.8% by engineering new features and building a fine-tuned ensemble of XGBoost models. Deployed AI models using AWS (S3, DynamoDB, Lambda, EFS, EC2, VPC, Cognito) and integrated them with the frontend. Built a lightweight pipeline to retrieve data from DynamoDB for m