Final undergraduate research project at University of Brasilia applying DINOv3 Vision Transformers for unsupervised species-level semantic segmentation of aerial vegetation imagery. Achieved 8.7% mIoU (52% of supervised ceiling) using DINOv3-Large with GMM clustering and PCA dimensionality reduction. Target application: biodiversity monitoring and species richness estimation from aerial surveys. Built custom experiment tracking system (MetaStore) with content-addressable storage, hardware-aware ETAs, config hashing, and auto-tagging. Enabled 100+ systematic parameter sweeps with reproducible workflows and comprehensive evaluation on FORTRESS aerial imagery dataset.

A declarative, reproducible terminal development environment delivered via Determinate Nix and devenv. Solves Helix/Zellij keybind conflicts through smart rebinding. Zero-config setup with intelligent pane management, sidebar reveal, and first-class Helix/Neovim support. Blazingly fast launches (~0.5s) with identical environments locally and over SSH. The name comes from Yazi + Zellij + Helix.

Actively maintained Yazi plugin that automatically adjusts the file manager layout based on terminal size. Provides smart column management (parent, current, preview) for optimal viewing experience, especially useful in sidebar configurations. Updated to support recent Yazi versions with thoughtful error checking and logging improvements.