Problem: Driver fatigue and cognitive overload are major safety risks in autonomous and semi-autonomous driving. Relying on a single sensor isn’t enough to catch it reliably.

Approach: Built a multimodal AI pipeline combining visual signals (eye and mouth activity, pupil features) with physiological and vehicle-context data, using temporal deep learning (CNN-LSTM) to detect fatigue patterns over time rather than frame-by-frame.

Outcome: A research-grade driver monitoring pipeline used in real-world experimental sessions. Contributed to applied autonomous-vehicle safety research and publication outputs.

Python PyTorch/TensorFlow OpenCV CNN-LSTM RealSense CAN bus

Problem: Autonomous-systems research needs reliable, synchronized data from many different sensors, without losing timing consistency or data quality.

Approach: Designed a ROS2-based data acquisition workflow integrating cameras, stereo vision, thermal imaging, LiDAR, GNSS/IMU, CAN bus, and physiological sensors for repeatable real-world data collection.

Outcome: A structured multimodal data pipeline that improved repeatability and sensor organization, supporting autonomous systems experiments.

ROS2 Python OpenCV ZED stereo LiDAR NVIDIA Jetson

Problem: UAV imagery is hard to analyze automatically. Altitude, angle, lighting, and clutter all change constantly, and labeled data is limited.

Approach: Built computer vision pipelines for detection, segmentation, and analytics from aerial imagery, including preprocessing, model training/evaluation, and real-world field validation.

Outcome: UAV-based AI workflows for real-world monitoring and inspection tasks, moving from research prototype toward operational use.

Python PyTorch TensorFlow OpenCV NVIDIA Jetson

Problem: Autonomous systems need to understand dynamic scenes, objects, lanes, road context, in real time and under real-world conditions.

Approach: Developed a visual scene analysis pipeline combining object detection, tracking, and segmentation for real-time processing on autonomous vehicle research platforms.

Outcome: A modular scene-understanding and tracking pipeline used to evaluate AI behavior under motion, viewpoint changes, and sensor noise.

Python OpenCV YOLO-style detection ROS2 ZED stereo

Problem: Industrial and semiconductor imaging requires reliably detecting fine visual defects and quality issues from high-resolution images.

Approach: Developed image-processing and AI pipelines covering acquisition, preprocessing, feature extraction, and defect/anomaly analysis, built for production reliability.

Outcome: Contributed to production-grade computer vision systems for industrial inspection and metrology.

Python C/C++ OpenCV Defect detection Git

Problem: Roads develop potholes that are costly and time-consuming to survey manually, and identifying both location and true depth is difficult at scale.

Approach: Used a drone with a professional camera and a 64-channel Velodyne LiDAR to collect imagery and laser scan data, then trained a machine learning model for pothole detection and depth estimation on a GPU server.

Outcome: The model successfully detected potholes and measured depth with strong accuracy, a scalable alternative to manual road-condition surveys.

Python PyTorch/TensorFlow LiDAR UAV imagery GPU training