metadata
title: Agentic Reliability Framework
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: static
pinned: false
license: apache-2.0
short_description: Bayesian governance for agentic systems
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