Shameemudheen Edayattu

• Engineer the multi-sensor perception pipeline for an autonomous rover's hitch, travel, and docking system, integrating 4 Ouster LiDARs (32- and 128-layer) and wide-angle/standard cameras with the Emergency Braking System
• Redesigned the UDP packet interface for the rover perception system and introduced a new trailer_id field to support trailer identification and tracking
• Improved keypoint and bounding-box detection on a custom Detectron2 model by moving to a ResNet-18 backbone and applying transfer learning via layer freezing; resolved the resulting catastrophic forgetting using Elastic Weight Consolidation for stable, incremental updates
• Diagnosed and resolved a Docker-related bug affecting ROS bag storage within the Emergency Braking System
• Contributing to an AWS-based MLOps pipeline and YOLOv8 segmentation model for a multi-camera railcar asset-tracking system, from data annotation through Jetson edge deployment

• Architected a production GenAI-based rail coupler defect-classification system for client using conditional GANs (SPADE) to generate photorealistic synthetic training data for a defect type with zero real examples; deployed on Azure Cloud with a live analytics dashboard, reaching 95% classification accuracy in production — Project Excellence Award
• Led the client off-road drivable-path segmentation project end-to-end from proof-of-concept through client delivery: built a PointNet++ 3D segmentation model on stereo-camera point cloud data, deployed to an NVIDIA Jetson Orin edge device, and applied hierarchical clustering to rank drivable paths; managed a team of 5 annotation engineers — Project Excellence & Extra Mile Awards
• Built an in-house C++17 image-enhancement pipeline using PhaseOne's ImageSDK to replace a manual, third-party process for client's drone-captured railway imagery, containerised with Docker for edge/cloud deployment — recognised with Tata's group-wide Innovesta Award (Best Innovative Project, all Tata companies)
• Led a team of 6 engineers and 6 interns to build an excavator missing-tooth inspection system for client, combining SAM2 and DINO segmentation with Blender-generated synthetic data to overcome a total lack of real defect imagery — Extra Mile Award
• Delivered rail-gauge and track-gap detection (YOLOv5/v7/v8) as part of CSX's railway track health monitoring system, and automated face/license-plate anonymisation for a video-privacy compliance project using YOLOv7 and OpenCV
• Ran a live CAN-data-transfer demo between an NVIDIA Jetson Orin board and a laptop for a client, directly contributing to a new engagement staffed with 14 engineers

• Self-initiated and delivered a Mask R-CNN semantic segmentation system to identify wooden railway sleepers under
overlapping ballast coverage from drone imagery (90%+ test accuracy); published as a poster at IRSC 2023 and demoed
live on a miniature track model to Tata's Advanced Systems team
• Contributed to a YOLOv5/v7/v8-based railway track health monitoring system, including image preprocessing and
MongoDB-based metadata queries to classify drone imagery by view type