Yukti Kumar

Recently graduated Electronics & Communication Engineer (B.Tech, PES University) with hands-on experience building perception and localization stacks for off-road autonomous platforms. Developed a full SLAM, visual-inertial odometry and sensor-fusion pipeline for an autonomous agricultural tractor at Koyth Robotics, deployed on Jetson Xavier NX. Two published IEEE papers spanning physics-informed machine learning for analog circuit design and quantum-seeded cryptographic steganography.

Comfortable across the stack – from RealSense/GPS/IMU hardware interfacing in C/C++ to factor-graph optimization, containerized ML pipelines, and edge deployment.

Robotics & AI Intern - Koyth Robotics - Bangalore

(2025-01 - 2026-01)

• Built a modular monocular Visual-Inertial Odometry package (vio_pipeline) with GPS-scaled translation chaining for absolute-scale recovery and Madgwick AHRS IMU fusion; resolved a critical two-clock-domain mismatch between wall-clock and device-monotonic timestamps.
• Iterated a RealSense D435i visual odometry pipeline through 8 versions, engineering support for depth vs. non-depth bag formats, window/glass masking, floor-plane RANSAC segmentation, and dual-mode LK optical-flow + ORB feature tracking.
• Designed and deployed an outdoor SLAM stack for the Mahindra Jivo 4WD tractor utilizing LightGlue/SuperPoint feature matching, factor-graph pose optimization, loop closure, and 3D point-cloud reconstruction.
• Built a persistent global obstacle-mapping pipeline integrating SegFormer semantic segmentation, RANSAC ground-plane fitting, and ENU coordinate projection – resolving heading drift and depth artifacts via median-blob-depth anchoring and LinearStructureMemory.
• Engineered an end-to-end GPS/IMU acquisition system featuring binary AA55 packet framing, NMEA parsing, a custom Tkinter GUI telemetry recorder, and a live dead-reckoning trajectory plotter.
• Containerized Mask2Former + SegFormer-B5 segmentation pipelines into optimized Docker images for AWS ECR, establishing an automated annotation and training pipeline on AWS SageMaker with S3 integration.
• Deployed perception models to Jetson Xavier NX via TensorRT for real-time, low-power edge inference; engineered a terrain-compensation script resolving GPS-antenna/camera lever-arm offsets in the ISO 8855 body frame.
• Developed a grape-bunch detection and unique-counting system using Mask R-CNN on the WGISD dataset paired with monocular-odometry-based bird's-eye-view (BEV) tracking.

B.Tech. - Electronics & Communication Engineering - PES University (2022-01 - 2026-01)

Pre-University (PCMC) - Deeksha (2020-01 - 2022-01)