Skip to content

INTYGA Intelligence · How it works

From a connected signal to an informed decision.

Intelligence collects the activity you connect, looks for specific conditions and keeps the supporting evidence alongside each finding. You decide whether it needs a response.

  1. 01Connect

    Choose a source and the service it describes.

  2. 02Detect

    Apply rules to observed events and measurements.

  3. 03Inspect

    Read the evidence and its limits.

  4. 04Decide

    Investigate, record context or mark it expected.

Start with a service, not every account.

Create a service such as “Production API” and assign its related sources. A deployment and an error signal need to belong to the same service before they can support a regression finding. Separate connections keep unrelated environments apart.

Some sources send authenticated webhooks. Others are polled by the scheduler. Activation requires approval; creating a connector does not begin collection. The connector view distinguishes setup, waiting for data, healthy collection and failures.

See what each integration collects →

Rules look for specific conditions.

Authorization patterns
Refusal bursts against a subject's own history, with unfamiliar action types as context. A denial does not prove successful access.
Changes and delivery
Regressions following production deployments, repeated or aged CI failures, and accepted bypasses of failed active merge controls. Correlation does not prove cause.
Operational health
Repeated collection failures, overdue jobs, degradation against configured service objectives, and known INTYGA health metrics. Missing telemetry is not proof of an outage.
Provider alerts
Selected active cloud, monitoring, Workspace and Stripe alerts. The provider's condition remains identified as such; only an explicit matching recovery closes it automatically.

Targets and settings matter. CI checks need selected repositories, branches and workflows. Job checks need expected completion intervals. Service-level checks need objectives and enough complete measurements. A connected source does not make every detector applicable.

Know what the evidence actually says.

A detector finding

A measured condition crossed a rule: for example, refusals increased relative to observed history. Counts, windows and available source references explain why.

A provider alert

A connected provider reported a condition. Its severity is useful context, but does not independently prove an anomaly, compromise or customer impact.

A review observation

A change touched a sensitive code path. It appears under Observations; a path match or routine PR merge alone does not trigger an anomaly assessment or alert.

A deploy-regression finding joins a successful production deployment to a subsequent increase in error or incident signals. It reports timing and the available connection to code changes. Root cause still needs investigation.

Leave a decision your team can use.

Filter by service, status, severity or category. Open the evidence, record a note and mark the finding as investigating, expected, a false positive or fixed. Repeated observations of a condition are grouped instead of becoming an unrelated list of alerts.

Optional alert rules send matching findings to Slack or a signed webhook, with severity thresholds and cooldowns. Review observations do not trigger those alerts. Intelligence does not execute a fix in the connected provider.

AI is an optional second opinion.

The detectors work without a model. You can configure TypeSafe AI (Jev), OpenAI or Anthropic as a review provider. Once you approve turning reviews on, eligible findings can receive an advisory review based on a restricted projection of their evidence. You can also ask another configured provider for a second opinion after inspecting the exact request.

Guidance reflects the evidence group: code and delivery, identity, cloud operations, monitoring, payments or other sources. A routine merge is expected activity; a refused request can mean a control worked; a dispute can need attention without being unusual.

Anomaly likelihood, human attention and severity are separate questions. Missing context stays uncertain. A review never creates a finding or changes its severity, status or alerts, and it cannot authorize a change. Prior opinions are not fed back as evidence. Operational and native-provider detectors do not request model reviews.

Your deployment, with explicit data flows.

Stored with you
The application and PostgreSQL database run on your infrastructure. Connector credentials are encrypted at rest. Retained webhook payloads can contain provider data, so choose permissions and retention for the sources you connect.
Sent when configured
Provider polling uses the credentials you supply. INTYGA handles sign-in and protected-action approvals. Optional AI review sends projected evidence to the provider you configure: TypeSafe AI, OpenAI or Anthropic. Alert destinations receive their configured finding payloads.
Protected by approval
Activating a source, changing detector settings, enabling alerts, configuring judgment or purging data requires a signed approval for that change. Workspace rules determine who must approve. A passkey response alone does not mean the operation has been applied.
Bounded visibility
Intelligence sees only the connected sources and successfully collected events. Quiet webhooks, missing history or a failed poll do not establish that the underlying service is healthy. Its availability does not control INTYGA's authorization service.

Prepare a deployment and first connection →