X‑RoadAI
An AI layer for X-Road

Secure data exchange that now watches, predicts, and explains itself.

X-Road moves data between organisations with end-to-end integrity and non-repudiation. This edition adds an opt-in intelligence layer that runs beside the system — never in the signing path.

04
Capability pillars
44
Offline tests, green
00
Payloads ever read
100%
Advisory, non-blocking
01 / The operations loop

From raw signal to operator action — on metadata alone.

Four stages, executed by a read-only sidecar that never reads message contents.

01
Detect

Anomaly detection

A federated temporal call-graph model flags traffic, latency, fault and fan-out anomalies per edge.

02
Correlate

Incident clustering

Collapses a wall of edge alerts into a few incidents, centred on the prime-suspect subsystem.

03
Predict

Breach forecasting

Projects each edge's trend to warn — with a confidence score — before a metric actually breaches.

04
Explain

Grounded copilot

Turns an incident into a plain-language, source-cited briefing with read-only next checks.

02 / What is inside

Additive, opt-in modules. X-Road runs normally with the AI off.

Every claim below maps to code in the repository — ports, endpoints and providers included.

ModuleCapabilitySurfaceStatus
ai-monitor Explainable Z-score + linear-trend anomaly detection over X-Road traffic. REST /api/v1/anomalies · :4200 Live
ai-insights Error, usage, capacity and security analysis — plus natural-language queries. REST /api/v1/insights · :4201 Live
common-ai Provider-agnostic core — Anthropic, OpenAI, or on-prem Ollama — with anonymiser, rate-limit and health checks. Library · Spring Live
xroad-ai-platform Federated graph detection, incident correlation, forecasting and grounded copilot. Read-only sidecar Live
admin-ui AI Monitoring section — anomaly dashboard, drill-down detail, provider & privacy settings. Security Server · Vue 3 Live
rust-accel Optional high-performance proxy, crypto and scanner components via gRPC / FFI. Proposal Roadmap
03 / Demonstration

The whole loop, offline, in milliseconds.

No live system, no network, no install. The core trains on synthetic traffic and reports.

$ make clean && make eval
window '2026-06-09': 95 calls, 4 edges, 5 nodes (baseline windows: 6)

4 ANOMALY ALERT(S)
  [  high] LATENCY_SPIKE  analytics → registry  — mean 2500ms vs baseline 119ms (z=569.2)
  [  high] VOLUME_SPIKE   client-a → registry   — 70 calls vs baseline mean 11.8 (z=69.6)
  [medium] NEW_EDGE       scanner → registry    — previously unseen call edge with 5 calls
  [  high] FAULT_SPIKE    client-a → payments   — fault rate 70% vs baseline 0%

3 CORRELATED INCIDENT(S)
  INC-1 [  high] registry (service) — 2 alerts [LATENCY_SPIKE, VOLUME_SPIKE] from 2 callers
  INC-2 [  high] payments (service) — 1 alert  [FAULT_SPIKE]
  INC-3 [medium] scanner (client)  — 1 alert  [NEW_EDGE]

1 PREDICTED BREACH(ES)
  [medium] latency  analytics → payments  — ~3.0 windows to breach (confidence 1.0)

READ-ONLY: PASS   NO-PAYLOAD: PASS   NFR: PASS   44 passed   ALL GATES GREEN
Fig. 1make eval, run offline with the Python standard library only.
04 / Safeguards

Government-grade guardrails, enforced in code and CI.

Not convention — gates. The build fails if any of these is violated.

  1. 01

    Opt-in by default

    Every AI feature is off until an administrator turns it on.

  2. 02

    Air-gap friendly

    Run entirely on-premise with Ollama — no cloud dependency.

  3. 03

    No payloads, ever

    AI consumes operational metadata only; the signing and message-log path is never read.

  4. 04

    Anonymised before egress

    Data is sanitised before any external call; nothing leaves without consent.

  5. 05

    Advisory, non-blocking

    Models alert, correlate, forecast and explain — they never block, route, or sign.

  6. 06

    Federated & local

    Models train per server; only aggregate deltas are shared, never raw data.

Read the source.

X-Road® is a registered trademark of the Nordic Institute for Interoperability Solutions (NIIS). The AI capabilities shown here are an additive, opt-in extension licensed under the MIT License. This is a community fork — not an official NIIS product page.

Built as one self-contained HTML file. No frameworks, no trackers, no build step. Set in Georgia & monospace.