Shameemudheen Edayattu

▸ ADAS Perception System: Optimized an AI-based autonomous driving rover perception solution consisting of camera,
LiDAR, and RADAR sensors for the emergency braking system, improving real-time object detection accuracy through multi
modal data integration.
▸ Sensor Integration & Pipeline Design: Configured and integrated Ouster LiDARs (32/128 layer), wide/normal cameras,
and switch networking into the emergency braking pipeline; designed and implemented interface modifications between
the rover perception system and the emergency braking system for production readiness.
▸ Machine Learning Model Optimization: Improved keypoints and bounding box detection on a custom Detectron2 model
by modifying the backbone to ResNet-18 and applying transfer learning (layer freezing, optimizer tuning, hyperparameter
optimization via TensorBoard loss analysis), achieving measurable detection performance gains.

▸ Generative AI - Coupler Inspection: Developed a production-ready GenAI-based coupler inspection system using domain
adaptive models and conditional GANs for photorealistic data generation; deployed on Azure Cloud with integrated
dashboards, achieving 95% accuracy in defect classification and significantly reducing human error in rail car safety
inspections.
▸ Computer Vision - Teeth Inspection: Engineered a vision-based excavator teeth inspection solution using foundational
models (SAM, DINOv2, YOLOv8) and ORB feature extraction, replacing manual inspection processes and improving
operational efficiency by 60%. ML Pipeline Perception: ML-based perception module drivable navigation point cloud data stereo cameras. full pipeline: data preprocessing,
feature engineering, model training, hyperparameter tuning, and post-processing. Deployed on NVIDIA Jetson AGX Orin,
reducing inference latency by 25%. MLOps & Optimization: high-performance segmentation inference module in C++ for YOLOv7, enhancing
processing speed by 20% through computational optimization and memory-efficient implementation. Production Deployment: image enhancement pipeline in C++17 PhaseOnes
ImageSDK for raw IIQ to TIFF/JPEG conversion. via Docker for scalable edge/cloud processing across varied bit
rates, demonstrating production ML deployment expertise. Predictive Analytics - Rail Safety: a computer vision solution for static-gauge measurement and gap detection using
YOLOv5, improving rail track inspection accuracy by 20% and contributing to rail safety innovations presented at IRSC 2023. Team Leadership & Mentoring: junior engineers on ML best practices, code review standards, and deployment
workflows; contributed to team decision-making on architecture and technology choices across multiple projects.

Machine Learning - Rail Inspection: Developed an ML solution for rail wooden sleeper inspection using Mask R-CNN
semantic segmentation; trained, evaluated, and optimized inference time and precision on GPU machines, demonstrating
strong foundational skills in machine learning model development and performance tuning.