Initial Offerings

Signal Finding

  • Fast nonlinear dependency discovery and feature scoring. Identifies predictive relationships that may be weak or invisible to linear correlation, including local, nonlinear, and higher-order structure.

  • Designed for rapid screening at scale. Fully parallelizable. Very competitive with substantially heavier dependence-testing and feature-selection methods in our benchmark work, while requiring a fraction of the computation—making repeated testing across large feature libraries practical.

  • More than correlation or mutual information. Probes dependence structure directly and can distinguish features with similar conventional summary statistics but materially different relationships to the target.

  • Well suited to continuous, mixed, and high-dimensional problems. Provides a lightweight way to rank, select, and prioritize candidate inputs before more expensive modeling.

  • Estimated Release - Q4 2026

Forecasting

  • Nonlinear representation, simple readout. Complex transformations are performed in explicit analytical coordinates, allowing forecasting with ridge regression or other lightweight models. Inputs, coefficients, and incremental contributions can be inspected directly.

  • Fast training, retraining, and inference. Typically hundreds or thousands of fitted parameters rather than millions. No GPU required. Near-instantaneous inference.

  • Captures both state and motion. Compact nonlinear coordinates, velocities, and multi-timescale state changes capture trend, periodicity, autoregression, nonlinear response, regime movement, chaotic and other complex dynamics.

  • Edge- and bandwidth-friendly. Predictions can be computed close to the data source and transmitted or stored as compact analytical telemetry rather than requiring continuous movement of high-resolution raw data.

  • Uncertainty quantification is integral to the architecture. Supports horizon-specific calibration and conformal prediction so forecasts can be accompanied by empirically qualified uncertainty rather than point estimates alone.

  • More transparent than black-box forecasting when interpretation matters. Architecture preserves substantially more visibility than large end-to-end deep learning models into what information is entering the forecast and how it is being used.

  • Estimated Release - 2027

Systems Monitoring

  • Decomposes what changed—not just whether something changed. Separates system movement into location, scale, marginal-shape, and dependency/copula components, with complementary views of relational and temporal structure. These structures are native outputs of the analysis, not explanations inferred after training.

  • Compresses very large systems into an analytical state. Converts hundreds or thousands of interacting streams into a compact, domain-agnostic representation of system position, coordinated motion, and state kinematics.

  • From alert to diagnosis. Combines anomaly/dependence detection with localization, historical context, and change attribution to reduce alarm noise and help identify which variables and relationships deserve attention.

  • Built for continuous monitoring and forecasting. The same compact state used for observation can be tracked through time, compared with historical states, and ultimately used to forecast state transitions and emerging regimes.

  • Lightweight and deployable throughout the stack. Can operate at the edge for local relational assessment, on-site for system-level monitoring, or centrally for fleet, functional, or organizational views—generally without GPU infrastructure.

  • Estimated Release - 2027