One AI-native engine for every high-stakes decision.
Ingestion to decision in one system, with every model, policy and AI call governed in one place. Built on open standards, so your models, features and data stay portable. It runs managed by us, in your cloud, or in your data center.
Command Center · Insights · Workbench: triage, briefs, rules, pipelines. The surfaces your team runs.
Rules + ML + severity bands: synchronous decisions, <100ms target†, every verdict explained and audited.
Training, registry, serving: versioned models on point-in-time-correct features. No shared black box.
API · Kafka · DB · file · webhook: schema auto-detect, canonicalization, PII masking, DLQ.
Six primitives. Every decision pack composes the same ones.
Nothing domain-specific is welded into the kernel. That is what lets fraud management become credit risk or underwriting without re-platforming, and why an agent, a principal and a mandate enter as canonical entities rather than as a separate product.
Multi-source and async, landing raw data with schema auto-detect, canonicalization, PII masking and a dead-letter queue.
Point-in-time-correct offline features for training, online serving for the real-time path. One definition, both worlds.
Versioned pipelines and models, hot-reloaded into serving. Your models, trained on your transactions.
Natural-language authoring, versioning, precision tracking and what-if simulation over your real history.
Queues, severity, SLAs and patterns: generic primitives every pack maps onto its own workflow.
Every decision emits one durable event and downstream actions subscribe by declared contract. A transactional outbox, not fire-and-forget.
A pack declares the entities, mappings, features, models, policies, outcomes and actions a domain needs, so specialization is data rather than a fork. See what each pack decides →
Heavy AI work never blocks the decision.
A synchronous engine scores every transaction in real time. A deep-triage agent (in build) investigates asynchronously, in parallel, and never in the authorization path.
One POST call in your transaction flow. The response carries the action, score, severity band, the rules that fired and the audit reference.
Related cases, device history and ring checks, assembled before an analyst opens the case. Shadow-tested before it ever enforces.
// response · 87ms
{
"action": "review",
"score": 0.91,
"severity_band": "high",
"reason": "velocity + geo mismatch",
"evaluators": [
{"rule": "cnp_velocity", "hit": true},
{"model": "txnCBM@v14", "score": 0.91}
],
"audit_ref": "dec_EXAMPLE…",
"latency_ms": 87
}
A hard timeout budget and a fallback verdict when it is breached. Fail-open or fail-closed is your policy, set in configuration, not ours.
<100ms† is a design target for the authorization path, validated on your traffic including tail latency. Not an average over a demo.
Deployment footprint, HA and recovery posture and the managed-update model are documented in the due-diligence pack, on request, before any pilot.
Autonomy you set, one rung at a time.
Each agent runs at the level you choose and rises only as it earns trust on your data. Every action is ledgered, anything irreversible or adverse stays human, and any level rolls back in one click.
Runs on live traffic, acts on nothing. You compare its calls to your team's.
Proposes the action with its evidence. A person decides.
Reversible actions inside guardrails you set: caps, rate limits, auto-pause. Every action opens a mandatory human review.
Acts within your caps, still bounded to reversible actions. Anything irreversible or adverse stays human.
The synchronous verdict stays in the engine; these agents never enter the authorization path. Designed for automated-decision rules (EU AI Act; GDPR Article 22): every automated action spawns a human review with a decision clock. The agent path is in build, recommendation-only and shadow-tested today, and the higher rungs are the designed model, to be proven with design partners.
Three ways to run it. One codebase.
The engine, the API, the packs and the audit trail are identical in all three. What changes is who operates the infrastructure and where the data sits.
Single-tenant per customer, in the cloud region you choose. Updates, monitoring and backups are ours.
SINGLE-TENANT · REGION OF YOUR CHOICE
Your own AWS, Azure or GCP account, inside your VPC and key management. We ship the release and support it.
YOUR ACCOUNT · YOUR KEYS · YOUR RESIDENCY
On-premises inside your perimeter, where the regulator or your policy requires it. Kubernetes on your hardware; a decision needs no egress.
IN-PERIMETER · NO BULK EXTRACTION
In every model, models train on your data only, the decision log stays queryable by you, and the exit path is documented before a pilot starts. Footprint, HA and recovery posture per model are in the due-diligence pack.
Governed AI is the platform, not a feature.
Every AI call routes through one governed pipeline: PII and injection guardrails, spend budgets, circuit-breakers, full logging.
Keycloak / OIDC SSO and role-based access on every surface and API. Future agents run as named, audited identities.
Idempotent decisions, durable decision events and replay: an evidence trail structured for regulatory reporting.
SAMA CFF and CBUAE 3057 control mappings are written. Mappings for other regimes are produced with the first partner in that jurisdiction.
PostgreSQL, Redis, Kafka, Spark, MLflow, pgvector, KServe, Keycloak, OpenTelemetry. Portable by construction.
Score, features, rule hits and model version attached to every decision, readable by analysts, auditors and regulators.
We hold the code that makes your decisions to the posture we ask you to hold us to for the decisions themselves. Nyx is built with NOX, our own development harness: one ticket in, one reviewed pull request out, with scope planned up front, review and security gates that cannot be skipped, and a recorded trail for every change. NOX is in private beta in the NaxVerse tools library.
Open the NaxVerse tools library ↗