ARF Logo
Pilot‑first AI reliability Self-healing systems

Agentic Reliability
Framework

AI Reliability & Self-Healing Control Plane.

Turn probabilistic AI into deterministic, auditable action. Reduce MTTR with automated detection, decisioning, and recovery.

Outcome
Up to 85%
MTTR reduction
By shortening the detection-to-recovery loop.
Operating model
Observe
Decide
Heal
A control plane, not another dashboard.
What ARF is
A reliability layer for autonomous AI
Designed for agents, services, and observability signals in production.
Why it matters
Failure should be legible
If systems can act, they also need a way to explain and recover.
Deep tech, but operational: every signal must support a decision.
Problem Silent failure

Most AI systems fail silently in production.

  • Agents can drift without a clear failure event.
  • Decision pathways are often invisible once actions are taken.
  • Failures cascade before anyone notices the root cause.
  • Reactive monitoring sees symptoms, not intent or risk.
The real problem is not missing data. The problem is missing interpretation of the data you already have.
Failure mode: uncertainty becomes operational debt.
Consequence: debugging is slow, expensive, and incomplete.
Gap: standard observability does not equal control.
Solution Control plane

ARF turns probabilistic AI into deterministic, auditable action.

  • Real-time anomaly detection across AI agents and services.
  • Automated recovery actions when risk exceeds policy thresholds.
  • Audit trails for every decision, trigger, and outcome.
  • A feedback loop that improves future interventions.
Core loop
Detect → Decide → Act
Close the loop before failure propagates across the system.
Design goal
Make recovery explicit
The system should show why it acted, not just that it acted.
How it works System flow

ARF core engine: from signals to recovery actions.

Input sources
Agents / Services
Execution events from production workloads.
→
Observability
Metrics / Logs
Telemetry normalized into a common stream.
→
Interpreter
Bayesian Risk Engine
Scores uncertainty, severity, and propagation risk.
→
Action
Healing Intent Engine
Recommends and triggers recovery policies.
risk ≈ uncertainty × impact × propagation × time_to_detection
The point is not perfect prediction. The point is earlier, better-justified intervention.
Try it now Sandbox API

Test the reliability loop with our sandbox endpoint.

Example sandbox request (mock response): 
curl -X POST https://sandbox.arf.dev/v1/evaluate \
  -H "Content-Type: application/json" \
  -d '{
    "service_name":"api",
    "event_type":"latency",
    "severity":"high",
    "metrics":{"latency_ms":450}
  }'
⚠️ This sandbox returns mock data only. The real ARF engine is access‑controlled.
  • Returns a conceptual HealingIntent example.
  • Illustrates risk score and recommended action structure.
  • Designed to show the API contract, not production inference.
  • For full engine access, request a pilot.
Observability becomes operational control – with the protected engine.
Ecosystem overview Composable platform

ARF is a layered platform, not a single model.

🔬 Research

Mathematical foundations for hybrid Bayesian inference, risk modeling, and policy decisions.

🛠️ Protected Core Engine

Bayesian risk scoring, semantic memory, governance loop – access‑controlled, pilot only.

⚡ API Control Plane

FastAPI service exposing incident evaluation, feedback, and memory endpoints (gated).

💻 Frontend UI

Next.js dashboard for visualizing risk, actions, and system state (public demo UI).

🏢 Enterprise

Outcome‑based pricing, unlimited evaluations, SSO, SLA, on‑prem deployment, dedicated support.

🎯 Outcome

A single control surface for reliability across autonomous systems.

API surface OAS 3.1

Core endpoints expose the reliability loop directly.

GET /
Root. Basic service entry point.
GET /health
Health check for deployment and uptime validation.
GET /v1/get_risk
Returns risk estimation for a service or event context.
GET /v1/history
Fetches past incidents, outcomes, and decisions.
POST /v1/incidents/evaluate
Evaluates an incident and returns a HealingIntent.
POST /v1/feedback
Records whether the recommended action worked.
GET /v1/memory/stats
Shows memory statistics and retrieval state.
GET /openapi.json
Machine-readable API specification for integration.
⚠️ These endpoints are part of the protected engine and are not publicly accessible. Pilot customers receive API credentials.
Schemas include validation errors, incident requests, and structured output for integration and automation.
Traction Demo Call to action

Prototype deployed. Live demo ready. Built for early adopters.

  • Working prototype available through pilot access.
  • Documentation and public specification are on GitHub.
  • The system is designed for technical teams that need reliability, not just insights.
  • Next step: validate against real production failure modes with a pilot.
Landing Page GitHub Foundation Specification Sandbox API
What I’m looking for
Collaborators and early adopters
Teams building autonomous or agentic systems in production.
How to start
Scan the QR code to request pilot access, or visit arf-frontend-sandy.vercel.app/signup.
Pilot request QR

Request Pilot

GitHub QR

GitHub (Public)

ARF Logo
ARF — Agentic Reliability Framework Juan Petter Pilot‑first · Outcome‑based pricing
⚠️ Core engine is access‑controlled – not open source. Pilot only.