Kashaf Noor

Multi-agent LLM pipeline (Planner/Verifier/Inspector) for a simulated G1 humanoid pick-and-place assembly cell in MuJoCo, with RAG-grounded safety reasoning

Deep learning pipeline for brain tumor detection and segmentation on BraTS 2018 MRI scans — compares U-Net, a custom architecture, YOLOv8-Segmentation, and Mask R-CNN using Dice coefficient evaluation.

LLM-planned robot task execution with a RAG-grounded, explainable safety layer — benchmarks naive vs. speed-monitoring vs. speed+evasion collision avoidance (n=100, PyBullet) and finds speed-scaling alone isn't statistically safer, evasion is.

Benchmarks a cost-latency optimized RAG router: static vs adaptive top-k retrieval. Adaptive router uses similarity-score elbow detection to right-size context per query — no extra LLM calls. Tested on 50 QA pairs via Gemini: 17% fewer input tokens than static k=5, with higher Exact Match and F1. Includes a data-driven margin calibration script.

Surface defect detection on rolled steel coils/sheets, comparing YOLOv8/YOLOv11 supervised detection against PatchCore/PaDiM anomaly detection (via anomalib) for rare-defect scenarios like cracks, scale, and pitting. Trained on the NEU and Severstal datasets as part of an Arbotrix intern research track.

General-purpose visual defect detection for production lines, spotting surface flaws, misalignment, and missing components across metal, plastic, and cast parts. Combines YOLO-based localization with ResNet/EfficientNet classification, trained on MVTec AD and casting datasets, with a Gradio demo interface.