xandykati98

I build for the web and train models that learn. I care about the intersection of good engineering, hard problems, and novel ideas. Some projects:

SimpleCTM single-file PyTorch take on Sakana AI’s Continuous Thought Machines.

Here I learned a lot about research reproduction, PyTorch and it's quirks, and of course Continuous Thought Machines
perceiver PyTorch take on DeepMind's Perceiver (General Perception with Iterative Attention, 2021).

Here I learned a lot about training on the cloud, interpretability and the power of latent reasoning
3d-mnist real-time 3D dense-net MNIST visualizer; JAX, React, and Three.js.

On this I had a lot of fun with Three.js and refreshed on the fundamentals of neural nets and JAX
dfu-segmentation-net DFU binary segmentation (EfficientNet-B4 U-Net) with a secondary verifier to cut false positives.

Here I worked on diabetic foot ulcer segmentation: an EfficientNet-B4 U-Net (Dice ~89%, IoU ~81%) and a verification classifier on RGB plus mask to filter false positives from skin lesions and other hard negatives
PDPpredict; predict immunogenic response of pMHC complexes from PDB structural data and electrostatics.
py-docktope; DockTope-style pMHC pipeline (D1–EM–D2) in typed stages; main.py orchestrates PyMOL, GROMACS, Gnina, and packaging.