Omer Tariq

Leading 3D/4D scene perception for autonomous urban delivery robots, developing a production multi-camera pipeline unifying 3D object detection, monocular depth estimation, open-vocabulary recognition, and semantic segmentation. Proposed quantization and distillation frameworks for vision-language models and prototyped long-horizon vision-language-action policies.

Improved model robustness by 15% and inference throughput by 8x using diffusion-based data augmentation, self-supervised pretraining, and PyTorch-TensorRT optimization. Deployed a real-time PRPD fault-detection service on Azure Kubernetes Service with an MLOps stack.

Proposed domain-invariant architectures for inertial localization, a hardware-aware multiscale transformer for TinyML, and adapted BEVFormer for embedded multi-camera 3D detection. Fused multi-sensor streams and developed privacy-preserving federated learning systems for on-device odometry.

Integrated LiDAR, IMU, and wheel-odometry SLAM for an industrial autonomous mobile robot on power-optimized embedded hardware, meeting a warehouse positioning target of less than 5cm error.

Led SoC/RTL verification and FPGA development for embedded AI across four production programs, contributing to two first-pass silicon tape-outs.

Designed radiation-tolerant FPGA-based DSP systems for satellite communication and imaging payloads sustaining over 99.9% uptime in orbit; recognized with a Distinguished Service Award for the PakTes-1A launch.

Grade: 3.73/4.3

Activities and societies: KAIST MS/Ph.D. Scholarship

Ph.D. in Computer Science with a dissertation on robust domain-invariant inertial localization in real-time federated edge computing. GPA 3.73/4.3, advised by Prof. Dongsoo Han.

B.Sc. in Electrical Engineering with a thesis on real-time object detection and autonomous control using a 3-DoF robotic arm.