Teach a PyTorch model something new on a live stream, and it wipes out what it learned yesterday—unless you burn GPU hours replaying old data. Use XGBoost or Random Forest instead, and you can’t learn online at all; every new sample forces a full retrain from scratch.I wanted an engine that learns like a real-time control system—deterministic, single-pass, and immune to catastrophic forgetting—so I built Tech-RKJ (pip install tech-rkj).Underneath a clean, scikit-learn-compatible Python API (SubspaceClassifier and SubspaceRegressor), the entire math core is written from scratch in Rust with hand-tuned AVX2 + FMA SIMD kernels. Zero backpropagation loops. Zero epochs. Zero replay buffers. It solves optimal ridge-regularized subspace weights analytically in a single closed-form pass, and maintains a full inverse-covariance memory map ($P^{-1}$) that absorbs live streaming samples via Sherman-Morrison rank-1 updates in ~105 microseconds—writing new knowledge only into unoccupied subspace directions so old classes stay intact.We put it head-to-head against PyTorch and XGBoost across 8 multi-domain benchmarks, and the numbers speak for themselves:Deep Learning Continual Learning (Frozen ResNet-18 512-D on CIFAR-10): After streaming brand-new unseen classes (ship & truck) one sample at a time with zero replay buffer, a PyTorch Linear Probe collapsed from 85.95% down to 13.85% (72.10% forgotten). Tech-RKJ retained 83.67% (1.83% drop) while hitting 91.70% on the unseen classes.Tabular Classification (32-D, 15k rows with 12 pure-noise columns): Built-in supervised ANOVA F-statistic weighting suppressed the noise features automatically, hitting 84.90% accuracy in 4.77 MB RAM—beating both XGBoost (83.83%) and a PyTorch MLP (72.13%).Non-Stationary Concept Drift & Few-Shot Zero-Day Injection: Recovered from 18.07% → 82.20% after a sudden mid-stream polarity flip (where static XGBoost stayed stuck at 12.60%), and learned a brand-new zero-day signature from just 40 samples after 6,000 normal samples with 82.14% detection and 99.00% normal retention.Speed & Energy Efficiency: Completed full closed-form training in 43.39 ms (0.23 J)—24.8x faster and 15x more energy-efficient than an MLP (1,075.70 ms, 3.47 J).Pre-compiled universal manylinux wheels (Python 3.9–3.13) are live on PyPI—no Rust toolchain or GPU required:Bashpip install tech-rkj
Explore the Rust source, reproducible benchmark scripts, and API docs on GitHub: https://github.com/jay-123-oss/tech-rkj