🛡️ AegisAI: Production ML Trust & Reliability Shield
AegisAI is an open-source, model-agnostic, and domain-agnostic reliability shield designed to sit in front of any Machine Learning or Deep Learning model.
🌐 Live Interactive Demo: Try AegisAI in your browser
✨ Core Capabilities
- 🎯 Probability Calibration: Temperature scaling, Platt sigmoid, and non-parametric isotonic regression.
- ❓ Predictive Uncertainty: Normalized Shannon Entropy measures epistemic ambiguity and decision boundary risk.
- 🚨 Out-of-Distribution (OOD) Detection: Multivariate Z-Score distance profiling identifies adversarial or out-of-distribution inputs before inference damage occurs.
- 🛡️ Composite Trust Score ( - 100%$): Synthesizes confidence, uncertainty, and OOD penalties into an actionable reliability index.
- ⚖️ Automated Governance Decisions: Automatically assigns ACCEPT, CALIBRATE, HUMAN_IN_THE_LOOP_REVIEW, or REJECT.
📦 Quickstart & Usage
1. Installation
ash git clone https://github.com/arshavardhan/AegisAi.git cd AegisAi pip install -e .
2. Load and Predict with this Pre-Calibrated Bundle
`python from aegis import AegisModel
Load the model bundle directly
shield = AegisModel.load('model.aegis')
Predict with full reliability diagnostics
sample = [[0.3, -0.2, 0.4, 0.1, -0.3]] report = shield.predict(sample)
print(f'Prediction: {report.prediction}') print(f'Trust Score: {report.trust_score * 100:.1f}%') print(f'Governance Action: {report.recommendation.value}') print(f'Uncertainty: {report.uncertainty:.4f}') print(f'OOD Anomaly: {report.ood} (Z-Score: {report.ood_score:.2f})') `
3. Wrap Any Custom Estimator
`python from aegis import AegisModel from sklearn.ensemble import RandomForestClassifier
1. Train your estimator
clf = RandomForestClassifier().fit(X_train, y_train)
2. Wrap with Aegis
shield = AegisModel(clf) shield.fit(X_train) # Fits reference OOD distributions shield.calibrate(X_val, y_val, method='isotonic')
3. Save as portable bundle (< 150KB)
shield.save('my_model.aegis') `