@article{vyas2026fpgl,
  title   = {FPGL: A Meta-Learned Implicit Neural Representation Framework for Medical Image Segmentation},
  author  = {Vyas, Kushal and Kim, Daniel and Netherton, Tucker and Veeraraghavan, Ashok and Balakrishnan, Guha},
  journal = {Medical Image Analysis},
  volume  = {115},
  pages   = {104316},
  year    = {2026},
  doi     = {10.1016/j.media.2026.104316}
}

Meta-learns a shared INR initialization by jointly fitting scans and segmentation maps, so novel subjects can be segmented simply by fine-tuning on their raw test-time observations.

@inproceedings{vyas2026surprising,
  title     = {The Surprising Effectiveness of Noise Pretraining for Implicit Neural Representations},
  author    = {Vyas, Kushal and Kayabasi, Alper and Kim, Daniel and Saragadam, Vishwanath and Veeraraghavan, Ashok and Balakrishnan, Guha},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

Pretraining INRs with structured noise yields dramatic signal-fitting improvements. Noise with 1/|f^α| spectral structure matches the best data-driven initialization methods.

@article{kim2024oceanai,
  title   = {Engineered Underwater Vehicle for Ocean Litter Mapping},
  author  = {Kim, D.},
  journal = {New Mexico Journal of Science},
  year    = {2024}
}

An autonomous underwater vehicle using YOLOv5 and SLAM to map subsurface debris, identifying 174,000+ pieces of litter.

@article{kim2022tvc,
  title   = {Developing a Control Algorithm and Simulation for Thrust Vector Controlled Rockets},
  author  = {Kim, D. and Iturregui, A.},
  journal = {New Mexico Journal of Science},
  year    = {2022}
}

A PID-based control algorithm and 6-DOF flight simulation for model rockets with thrust vector control.

A photorealistic, differentiable drone simulator built from 3D Gaussian Splat reconstructions, used to train monocular RGB navigation policies across diverse scenes.

A custom Teensy-based flight computer and thrust vector control system that actively stabilizes model rockets during ascent, starting from PID control with Kalman filtering. Ember extends it with a differentiable 6-DOF rocket simulator that trains a neural TVC controller by backpropagating through the physics, then flies the learned policy on the real hardware.

A modular PyTorch library for rapidly prototyping implicit neural representations, with object-oriented interfaces and visualization tools for scientific computing.

An autonomous underwater glider with a custom YOLOv5 vision pipeline and SLAM navigation that identified over 174,000 debris objects.

A real-time productivity platform with focus tools, task sync, and Lo-Fi streaming, serving 1200+ users.

An N-body gravity simulator visualizing 8,000+ active satellite trajectories in real time, built on a custom C++ engine with Vulkan rendering.