Kazim Mohammed Abbas

• Help lead the MineOcclude research project, which studies how objects remain represented when they become hidden in controlled Minecraft scenes.
• Coordinate work across data collection, computer vision, geometric calibration, visibility labeling, testing, and research documentation.
• Built and finalized an HSV-based object detector for 640 × 360 images. It reached 100% precision and recall on a balanced 72-frame development dataset, with development data kept separate from the final test set.
• Designed a 48-frame held-out test set with 24 positive and 24 negative examples across six layouts, using fixed labels, camera positions, and capture conditions.
• Helped validate six homography models used to convert image locations into Minecraft world coordinates, reaching about 0.093 blocks median error and 0.287 blocks at the 95th percentile.
• Helped create a visibility dataset with 1,850 labeled records across 56 trajectories and six layouts.
• Built automated checks for data quality and reproducibility using Python, NumPy, pandas, PyArrow, pytest, Ruff, Git, SHA-256 hashes, JSON/JSONL, and Parquet.

Built an AI-based surveillance and situational awareness system using real-time computer vision to detect and surface events. Developed the frontend interface for displaying detected events and system outputs in a clear, organized workflow. Contributed to an agent-based routing system that directed detected signals to different response paths based on context. Tested system behavior, adjusted parameters, and documented limitations including false positives, edge cases, and performance tradeoffs.

Participated in Google DevFest, working on a computer vision prototype that simulated automated defect detection for infrastructure inspection. Built a Python application that processed image inputs, highlighted potential defects, and displayed confidence scores using a simple visual interface. Integrated an AI-based analysis step to evaluate detected defects and trigger alerts when issues exceeded a threshold. Focused on making system outputs easy to interpret, testing edge cases, and reasoning about false positives, reliability, and alerting logic.

Built a computer-vision–based adaptive traffic signal system using Python and OpenCV. Designed logic to detect vehicles from live video, compute rolling traffic demand, and dynamically adjust signal timing in real time. Implemented a state machine for green/yellow/red phases, integrated hardware light controls, and modeled fuel and CO₂ impact from reduced idle time. Focused on reliability, noise reduction, and real-time decisionmaking under uncertainty.