Keerthan HS

Designed and implemented an end-to-end ETL pipeline using PySpark and Databricks to ingest, clean, transform, and store raw CSV data in optimized Parquet format. Performed data annotation and labeling to generate bounding box annotations for supervised learning.

Implemented Medallion Architecture (Bronze, Silver, Gold), schema validation, incremental data loading, window functions, logging, and error handling to simulate a production-grade data engineering workflow. Developed analytics-ready datasets through data aggregation and transformation, improving query performance using partitioned Parquet storage.

Collected and curated pothole image dataset from Kaggle for model training Performed data annotation and labeling to generate bounding box annotations for supervised learning Preprocessed dataset and split into training and validation sets (80:20 ratio) Designed and configured custom YAML file to define dataset paths and class labels Trained an object detection model (e.g., YOLO) on annotated dataset Evaluated model performance using validation data and unseen test images Implemented inference pipeline to detect potholes and generate bounding box predictions on new images Visualized detection results for performance analysis and validation Delivered actionable insights and detection outputs to relevant government authorities for infrastructure monitoring

Managed the complete machine learning pipeline from data acquisition to model deployment in a production-like environment Collected and organized datasets from secure cloud storage (e.g., OneDrive) for multiple industrial components Performed image annotation and labeling for object detection and anomaly detection tasks Trained and optimized YOLO-based object detection models for identifying multiple component classes (e.g., rods, springs, mechanical parts) Implemented and evaluated deep learning models including ResNet50, ResNet101, and PaDiM for classification and anomaly detection Conducted model validation and testing on unseen data to ensure robustness and accuracy Built inference pipelines to generate predictions and integrate outputs into downstream systems Deployed trained models into an internal GUI-based application for real-time or batch predictions Collaborated in improving data quality, annotation consistency, and model performance across iterations Ensured adherence to data confidentiality and secure handling practices throughout the project

Secured 2nd prize in our mini project conducted in 6th semester