Designed and implemented an end-to-end evaluation pipeline covering data ingestion, ground-truth generation, simulation, and performance analysis.

Built a multi-process framework achieving 10× faster runtime, enabling large-scale validation.

Created a flexible data pipeline ensuring reproducibility and traceability across test runs.

Defined evaluation metrics and regression tests aligned with Euro NCAP, GSR, and ASPICE standards with automated reporting.

Developed debug-image generation and interactive visualization tools to speed up root-cause analysis and data-driven release decisions.

Automated Looker dashboards that update with each dataset and software release, visualizing key metrics—detection rate, tracking accuracy, positional and velocity errors—to monitor trends, detect false AEB (Automatic Emergency Braking) events, and guide release planning.

Designed a hardware-in-the-loop (HIL) framework validating perception software on TI AM62A and TDA4VH boards using real sensor replay.

Automated multi-board orchestration and YAML-based configuration for scalable parallel runs.

Built metadata-standardization tools for large rosbag datasets (~2 GB / 10 s), eliminating manual setup errors and improving efficiency.

Added validation checks (frame-rate, size, duration) with auto re-runs to ensure reliable ADAS results.

Built a semi-automated labeling pipeline replacing vendor and manual annotation with an in-house CVAT-based system.

Integrated Grounding-DINO detection and GRM tracking to auto-generate CVAT-compatible annotations.

Implemented label-schema version control and transitioned to teacher-model pseudo labeling, supporting 10 European countries and 416 traffic-sign variations.

Reduced manual labeling effort by > 90 % and enabled continuous updates aligned with new model releases.

Develped the method for the diversity-based image selection (diverse scenes / context) for deep learning model validation rather using the randomly selected data sets

Helped to have insights on rare / corner case, so later, we can collect more data for training

Selected mutual-exclusive images among the clusters; it

Completion of the two in-depth capstone projects

Extra Project: Data Anonymization for Privacy

De-identifying sensitive data by implementing anonymization techniques (e.g., generalization for bucketing, suppression, and pseudonymization)

Implemented the k-anonymity and the information loss for measuring privacy and accuracy of anonymized data

Observed the accuracy of machine learning models built with data of different levels of privacy (including differential privacy technique)

Received the two National Grants for my research projects, "Efficient Refactoring Candidate Identification" (Principal Investigator)

Published the research results in the top conferences and journals such as IEEE Transactions on Software Engineering

Fast graph-based coupling metric for calculating the effects of suggested refactorings based on matrix computation. This sacrifices some degree of precision (e.g., Connectivity: precision = 0.52, recall = 0.96 ) but significantly reduces the computation complexity, which can be helpful in analyzing large-scale software (e.g., our approach: 154 sec vs. EPM (previous approach): 28,622 sec).

Two-phased search-based refactoring identification method predicting the probabilities of the candidates that are likely to improve "maintainability" and evaluating only the candidates with higher chances. Compared to the previous research of the no-reduction approach, the two-phased refactoring identification approach is 2.6 (min) to 13.5 (max) times faster in computation time.