A self-contained KD-tree implementation for exact nearest-neighbor search in Nx tensors. It supports multiple distance metrics, leaf-size tuning, and query-time traversal. The module is isolated enough to be extracted into any Nx-based project needing efficient spatial indexing, and demonstrates careful numerical and recursive tree-building logic.

Implements the NN-descent algorithm for approximate nearest-neighbor search, a core building block for large-scale manifold learning and recommendation systems. It uses randomized projection forests and iterative graph refinement, all inside Nx defn loops for JIT compatibility. This is a non-trivial, reusable algorithm that would be difficult to reimplement correctly.

A full HDBSCAN implementation with mutual reachability, minimum spanning tree construction, cluster hierarchy extraction, and stability-based cluster selection. It handles numerical edge cases like infinities and overflow carefully. The module is a standalone clustering algorithm that can be reused in any Elixir/Nx data pipeline requiring robust density-based clustering.

A JIT-compatible BFGS quasi-Newton optimizer with line search, implemented as an Nx defn loop. It exposes a clean minimize/3 interface and is independent of any specific objective function, making it directly reusable for custom optimization problems in Elixir. The numerical care in handling gradients and step sizes demonstrates strong optimization craftsmanship.

Implements the Ledoit-Wolf shrinkage estimator for covariance matrices, a statistically sophisticated technique that improves conditioning when the number of samples is small relative to dimensions. The module is self-contained, exposes fit/transform-style APIs, and can be extracted for any statistical or financial application needing robust covariance estimation.