projects

A fully declarative Neovim setup powered by Nix and Lua. Batteries included, but minimal by default - tuned for speed, efficiency, and daily driver use.


Multi-agent RL benchmark supporting MADDPG, COMA and GNNs with dynamic agent adaptation.An open-environment multi-agent reinforcement learning benchmark suite with support for MADDPG, COMA, and GNN-based policies. Enables dynamic agent/task/frame adaptation for open-system evaluation.


Rust+CUDA backend with Python bindings that doubled benchmark performance.A Rust+CUDA backend to accelerate the free-range-zoo benchmark, doubling environment performance. Integrates Python bindings with Maturin to expose custom CUDA logic for seamless use in RL workflows.
Real-time 3D microscopy software processing 20GB/s using MPI/HDF5 and OpenCV.Real-time imaging software for a 3D 2-photon lightsheet microscope processing 20 GB/s of image data. Built with parallel MPI/HDF5 data acquisition and OpenCV pipelines for volumetric dataset processing.

Django interface with PyTorch models for agricultural root analysis (97% accuracy).Django interface for root segmentation and volume extrapolation. Built root segmentation models using PyTorch for high-resolution root image data, with OpenCV processing for volumetric estimates achieving 97% accuracy.

Computer vision system with 99% accuracy for counting overlapping fonio seeds.Compiled and labeled a dataset of over 8,000 fonio seed images for robust model training. Implemented a neural network achieving 99% accuracy in detecting overlapping seeds and providing precise counts.