Hi! I’m Kaustav. I’m currently an AI Engineer at Kaliber Labs in San Francisco, and a recent graduate of the CMU Robotics Institute, where I earned my Master's in Computer Vision.
My recent work spans from 3D vision and reconstruction pipelines to robotics and world models, with a sprinkle of AI conversational systems in between.
Introduced a generalizable training-time auxiliary supervision target for jointly-trained JEPA-style world models to better structure the latent space, improving performance across various benchmark environments and doubling success rate in low-data regimes.
Used truncated signed distance function (TSDF) fusion for real-time 3D reconstruction of the knee during arthroscopic surgery. We found that oblique viewing angles lead to significant drops in reconstruction quality, which has clinical relevance during surgery.
Used hierarchical Multi-Task Learning (MTL) and semi-supervised learning with Momentum Pseudo-Labeling (MPL) for phoneme recognition in non-canonical speech, achieveing SOTA phoneme-error rates (PER) on the L2-ARCTIC dataset.
Developed an autonomous end-to-end cinematographic trajectory generation pipeline with hollywood production company Camera Control.
Used bundle adjustment with neural image feature descriptors to improve gaussian splat 3D reconstruction quality from 360° images.
Created a pipeline to improve performance of SOTA 3D human avatar generation models on in-the-wild images.
Implementation and experimentation with neural plane optimization for 3D scene representation.
Master of Science in Computer Vision (MSCV)
Perception Research at CMU AirLab
Bachelor of Engineering in Mechanical Engineering
Highest Distinction · 2nd Major in Innovation and Design
Overseas Exchange Program
AI Engineer
Systems Engineering Intern
Robotics Engineering Intern
