SAE (State Aware Engine) is a cross-domain temporal intelligence framework for discovering, tracking, and reasoning about latent states and regimes in sequential data.
It is powered by NHSMM (Neural Hidden Semi-Markov Models) and designed to bridge research-grade sequence modeling with production-ready deployment.

SAE provides a unified engine for state inference, regime detection, and temporal decision support across heterogeneous, non-stationary time series.

• Explicit Latent State Modeling
  Models hidden regimes with variable, learnable state durations, enabling accurate dwell-time reasoning beyond standard HMMs.

• Context-Aware Dynamics
  Initial states, transitions, durations, and emissions can be modulated by external covariates, supporting non-stationary and adaptive behavior.

• Neural + Probabilistic Hybrid
  Combines classical HSMM structure with neural parameterization, preserving interpretability while increasing expressiveness.

• Exact and Approximate Inference
  Supports forward-backward likelihoods, Viterbi decoding, and differentiable training for neuralized latent states.

• Production-Oriented Design
  GPU-ready, batched, modular, and suitable for cloud, on-prem, or edge deployment.

• NHSMM
  A modular PyTorch library implementing neural, context-aware Hidden Semi-Markov Models, fully open-source.

• SAE
  A system-level engine built on NHSMM, providing:

  • domain adapters
  • inference pipelines
  • deployment patterns
  • cross-domain abstractions

SAE treats NHSMM as its latent state inference backbone.

• Finance & Trading — Market regime detection, volatility states, adaptive strategy modeling.
• Cybersecurity & Systems Monitoring — Hidden operational states, anomaly detection, behavior shifts in logs or telemetry.
• IoT & Industrial Analytics — Predictive maintenance, machine state monitoring, fault regime discovery.
• Health & Wearables — Activity segmentation, physiological state tracking, anomaly detection.
• Robotics & Autonomous Systems — Behavior monitoring, task phase detection, safety-critical state transitions.
• General Temporal AI Research — Neural HSMMs, hybrid probabilistic models, non-stationary sequence learning.

• Cloud / SaaS — Scalable, multi-tenant temporal analytics.
• On-Prem / Edge — Low-latency, privacy-preserving inference close to data sources.
• Accelerator-Ready — GPU-first execution with future support for additional backends.

⚠️ Alpha / Early Development

SAE is under active development.
APIs, abstractions, and deployment tooling may change before 1.0.0.

SAE currently relies on NHSMM as its core dependency:

pip install nhsmm
Higher-level SAE components, adapters, and services will be released incrementally.
Early-access versions and research previews are available via Patreon or subscription for controlled use.

SAE is released under a Proprietary License © 2026 AWA.SI.

The repository is public, but usage is restricted:

Viewing, cloning, and personal experimentation are permitted.

Redistribution, commercial use, or deployment without permission is prohibited.

Early-access releases may be provided to subscribers via Patreon, with copy usage governed by this Proprietary License.

For full license terms, see LICENSE.

NHSMM remains fully open-source (Apache 2.0), and can be used independently.

Development and research around SAE and NHSMM are supported via:

Patreon (early-access, research sketches, pre-releases)

GitHub Sponsors

Medium articles & research notes

See FUNDING.md for details on how contributions help sustain development.