Satyam Gaba

Pioneered the industry's first globally scalable 3D Digital Twin pipeline, integrating multi-sensory data and leveraging Blender/Open3D in Python to enable automated material assignment.

Designed and developed an ML pipeline for the "Segment Anything Mesh Model" to zero-shot label 3D meshes, optimizing material assignment and establishing a scalable framework for automated city-scale 3D model labeling.

Led a cross-functional team of 7 engineers in the end-to-end development of an autonomous tricycle cart, integrating Stereo Cameras, SLAM, and ROS for outdoor navigation and obstacle avoidance.

Engineered advanced neural rendering models (NeRF, Gaussian splatting) to enhance reconstruction quality of complex outdoor environments, integrating ICP for precise Point Cloud alignment within the Digital Twin.

Advised over 10 interns and capstone students on cutting-edge research projects focused on LLMs, Generative AI, and Model Optimization, fostering next-generation talent.

Developed CNN and Transformer-based ML models for 5G Channel Compression, leveraging PyTorch and Ray for distributed training, featured in Qualcomm's and Nokia's MWC Barcelona 2024 demo.

Optimized large ML models for Snapdragon phones, reducing memory computation from FP16 to UINT4 while maintaining accuracy through Post Training Quantization.

Spearheaded the development of a C-based ML driver framework, enabling model inference on Hexagon Tensor Processor with less than 10us latency overhead, showcased at MWC ‘23 and ‘24.

Constructed an ETL data pipeline using Dask on AWS, facilitating Exploratory Data Analysis on terabytes of 5G modem data from real-world deployments.

Devised "FluMo," a novel parameter using Modulated Spectrograms to quantify motion fluidity, improving SOTA ML algorithm accuracy by 0.2% on the PAMAP2 dataset.

Developed "LibIMU," a C++ library for synchronous data acquisition from multiple IMU devices at up to 100 Hz.

Developed a perception pipeline utilizing decision-level fusion, integrating classical CV and Deep Learning models (Faster RCNN, UNet, PSPNet) with PyTorch and OpenCV for rebar extraction from aerial images.

Utilized Point Cloud Registration to fuse Photogrammetry and LIDAR 3D data, achieving 0.92 precision and 0.98 recall with a 3 cm offset for 3D rebar detection.

Programmed a Qt-based GUI labeling framework in Python, reducing labeling time and effort by over 50%.