ML Models: Temporal Forecasting and Countermeasure Effectiveness Standard
PyTorch-based machine learning pipeline for predictive analytics including SLA forecasting, threat detection, and temporal spiking neural networks.
- Machine Learning
- PyTorch
- Prediction
- Temporal
What shipped
- CES Model: Countermeasure Effectiveness Standard with threat/sla/stability scores
- Temporal Forecaster: Spiking neural network for dynamic predictions
- Training pipeline: GridSearchCV hyperparameter tuning with Polars batch processing
- Model publishing: Hugging Face Hub integration (state_dict only, no pickle)
Model architecture
The CES model combines three signals:
- Threat score: Active incidents + high-latency trends
- SLA score: Calculated uptime against baseline
- Stability score: Inverse of incident count + latency pressure
The fourth model (Spiking Temporal Forecaster) provides a dynamic adjustment factor based on real-time telemetry patterns.
Production considerations
- Models train nightly via
train_all_modelscommand - State dict serialization prevents deserialization vulnerabilities
- Tenant models namespaced with hashed slugs on Hugging Face
- Async training via
ml_workerswith Redis/Dramatiq