Agentic Reliability Framework (ARF)
Open‑source advisory engine for cloud infrastructure governance.
ARF provides provably safe, mathematically grounded recommendations—approve, deny, or escalate—when users request provisioning, configuration, or access changes.
🔍 What We Do
- Bayesian Online Learning – Fast, conjugate updates for per‑category risk using beta‑binomial models.
- Offline Pattern Discovery – Hamiltonian Monte Carlo (HMC/NUTS) logistic regression captures complex interactions (time‑of‑day, user role, environment).
- Composable Policy Algebra – Build fine‑grained rules with AND/OR/NOT combinators.
- Semantic Memory – FAISS‑based retrieval of similar past incidents for context‑aware decisions.
- Deterministic Probability Thresholds (DPT) – Clear approve/deny/escalate decisions based on calibrated failure probabilities.
📊 Key Mathematical Insights
| Concept | Implementation |
|---|---|
| Conjugate Priors | Per‑category Beta priors, updated online with outcomes. |
| HMC Sampling | Logistic regression with NUTS, serialized to JSON for hot‑loading. |
| Risk Fusion | Dynamic weighted combination of conjugate, hyperprior, and HMC estimates. |
| DPT | Approve if P(failure) < 0.2; Deny if > 0.8; otherwise Escalate. |
🚀 Quick Links
- 📦 GitHub Repository – Source code, issues, contributions.
- 🤗 ARF v4 Demo Space – Interactive Bayesian risk dashboard.
- 🔧 Legacy API Demo – v3.3.9 API for backward compatibility.
- 📚 Documentation – Full specs, tutorials, and API reference.
🧪 Try It Now
Click the Spaces below to interact with live demos. The v4 Space showcases the full Bayesian engine with real‑time risk scoring and policy evaluation.
🤝 Contributing
ARF is open source under the Apache 2.0 license. We welcome contributions of all kinds—code, documentation, ideas, or feedback.
📫 Contact
- Email: petter2025us@outlook.com
- LinkedIn: petterjuan
- Book a call: Calendly