Research & Publications

Automated assessment in programming education typically reduces feedback to a single pass/fail signal, leaving students with little insight into code quality or reasoning. Autograder+ introduces a multi-faceted AI framework that pairs static code analysis — linting, complexity and test-coverage metrics — with semantic feedback from large language models, producing structured, rubric-aligned comments across correctness, style and design. The framework is built to scale formative feedback in large programming courses without sacrificing pedagogical depth.

Computer-vision and object-detection research is focused on improving human vision and the identification of objects in working scenarios. This research proposes an improved technique using the YOLOv5 model with an adaptive Kalman filter for real-time object tracking and estimation, addressing challenges such as camera alignment, image quality, darkness, occlusion and other environmental factors. The proposed technique achieves better performance by utilising mixed precision for processor selection and parallel distribution of features, demonstrating high model accuracy and real-time processing capability — a training accuracy of 98% and a testing accuracy of 95.4% for surveillance applications, with a throughput of 14.8 and an inference rate of 45 FPS at 1080p resolution. The work contributes to overcoming common tracking challenges and provides solutions for conditional processing cases.