What systems can Guardian integrate with?

Monitoring, MLOps, ticketing, review workflows, and data warehouses — via API, webhook, or batch ingestion.

Does Guardian replace existing MLOps tools?

No. Guardian sits above your stack and links existing signals into audit-ready evidence records.

What evidence can integrations capture?

Metrics, drift alerts, review notes, incidents, deployment lineage, and oversight activity — normalized into one dossier.

Do we need to replace our current monitoring tools?

No. Guardian is designed to sit on top of the monitoring and evaluation systems you already use. The goal is to add a governance and evidence layer, not to replace your existing stack.

Do we need to send raw model data?

Not for a first integration. A metrics-first approach can start from aggregate monitoring signals, model metadata, evaluation windows, and governance-relevant events.

Can we keep our own internal metric mapping logic?

Yes. Guardian can support both native metric ingestion and pre-translated ingestion, depending on how much control your team wants over the adapter layer.

What kinds of AI systems can Guardian support?

Guardian is designed for heterogeneous AI environments, including classification, regression, forecasting, computer vision, ranking systems, and LLM-based applications.

Is Guardian a legal compliance determination tool?

No. Guardian provides technical evidence infrastructure and governance workflows. It supports compliance readiness, but it does not replace legal assessment or formal conformity assessment.

Book a pilot discussion

Appendix · Source registry

Integrations

Connect the systems that already record AI activity. Guardian ingests signals from monitoring, MLOps, ticketing, review, incident, and oversight workflows, then links them into audit-ready evidence records.

Book a pilot discussion Map your evidence sources

No stack replacement required.

Source registry

Registered intake channels for production signals, review activity, and oversight events. Each source normalizes into the same evidence record — without replacing your existing stack.

Evidence intake registerREF · GRD-SRC-REG

Source ID	System	Channel	Signal type	Evidence role	Status
SRC-01	Python SDK	API	model.metrics	Ingest · production metrics	Linked
SRC-02	REST API	HTTP	review.event	Capture · review events	Linked
SRC-03	Webhooks	Push	incident.opened	Alert · incident intake	Linked
SRC-04	Hugging Face	Pipeline	model.version	Lineage · model registry	Active
SRC-05	Azure ML	Pipeline	deployment.log	Trace · deployment record	Active
SRC-06	AWS SageMaker	Pipeline	endpoint.event	Trace · endpoint activity	Active
SRC-07	Weights & Biases	Monitor	drift.metric	Signal · drift observation	Linked
SRC-08	Datadog	Monitor	alert.fired	Ops · runtime alert	Linked
SRC-09	Snowflake	Warehouse	batch.export	Archive · batch evidence	Available
SRC-10	Jira / ServiceNow	Ticketing	incident.ticket	Oversight · incident queue	Available
SRC-11	Internal GRC	Webhook	review.completed	Custody · sign-off trail	Available
SRC-12	Custom adapter	Batch	native.metrics	Map · team-controlled schema	Available

12 registered sources · normalized to evidence record · no stack replacement required

Evidence intake

How incoming signals become linked, reviewable evidence inside the dossier.

01
Capture event
Ingest the metric, alert, review note, or governance event from the source system — via API submission, inbound webhook, or scheduled batch pull.
02
Normalize payload
Guardian maps heterogeneous native metrics into a stable governance schema: drift scores, consistency checks, status changes, and evaluation windows.
03
Link to policy / custody
Attach the signal to the applicable policy frame, custody handler, and reviewer context so oversight activity stays traceable.
04
Attach to evidence record
Append the normalized output to the living dossier — timestamped, versioned, and ready for export when scrutiny arrives.

Intake flow

Existing monitoring and evaluation tools → Guardian API or webhook → Normalized governance schema → Evidence records, status tracking, and oversight workflows

Common questions

Integration essentials

What systems can Guardian integrate with?: Monitoring, MLOps, ticketing, review workflows, and data warehouses — via API, webhook, or batch ingestion.
Does Guardian replace existing MLOps tools?: No. Guardian sits above your stack and links existing signals into audit-ready evidence records.
What evidence can integrations capture?: Metrics, drift alerts, review notes, incidents, deployment lineage, and oversight activity — normalized into one dossier.

Two integration modes

Choose the intake path that best fits your engineering environment and internal control model.

Native metric ingestion

Send the raw or existing monitoring metrics you already compute. Guardian handles the translation into a normalized governance schema.

Example metrics

PSI
precision / recall
MAE / RMSE / MAPE
embedding drift
hallucination rate
NDCG / ranking metrics

Best for

Fast onboarding and pilot deployments

Pre-translated ingestion

If your team prefers to control the mapping logic internally, you can send Guardian-normalized fields directly.

Example fields

driftScore
fairnessScore
groupConsistencyScore
psi
disparateImpact
demographicParity
tprDifference

Best for

Teams that want tighter control over internal adapters and mapping rules

Built for heterogeneous AI environments

Different AI systems produce different native metrics. Guardian translates them into a common governance layer — standardize governance outputs, not model internals.

Model family	Native signals accepted
Binary classification	PSI, recall, disparate impact, demographic parity, equal opportunity gaps
Regression and scoring	MAE, RMSE, MAPE, calibration error, score-distribution drift
Forecasting	MAPE, SMAPE, forecast bias, residual drift
Computer vision	mAP, IoU, embedding drift, confidence drift, per-group recall
Ranking and recommenders	NDCG, MAP, MRR, CTR shift, exposure fairness
LLM and generative AI	Hallucination rate, groundedness, refusal rate, toxicity rate, task success, prompt drift

Evidence intake interface

API-based metric ingestion

Submit metrics from scheduled evaluations, batch monitoring jobs, or internal reporting pipelines. The API is an evidence intake interface — not a developer dashboard.

Typical payload includes

schema version
tenant or system ID
model type
run ID
timestamp and evaluation window
source metadata
native metrics
evaluation context where relevant

Illustrative intake payload

{
  "schemaVersion": "1.0",
  "tenantId": "example-enterprise",
  "systemId": "wafer-defect-detection-model-01",
  "systemName": "Wafer Defect Detection Model",
  "modelType": "computer_vision",
  "runId": "quality_monitoring_2026_05_12_1030",
  "timestamp": "2026-05-12T10:30:00Z",
  "window": {
    "start": "2026-05-12T09:00:00Z",
    "end": "2026-05-12T10:00:00Z"
  },
  "source": {
    "provider": "internal-monitoring",
    "pipeline": "manufacturing-quality-monitoring",
    "environment": "prod"
  },
  "nativeMetrics": {
    "mAP": 0.91,
    "embeddingDrift": 0.14,
    "confidencePSI": 0.18,
    "recall_groupA": 0.92,
    "recall_groupB": 0.87
  },
  "evaluationContext": {
    "groupType": "production_line",
    "reference": "line_A",
    "comparison": "line_B"
  }
}

Guardian stores the native metrics, applies the relevant mapping logic, updates governance state, and appends the result to the evidence trail.

Request API overview

Event intake

Webhook-based event integration

Use webhooks when governance-relevant events should trigger ingestion or downstream workflows automatically.

Inbound webhook examples

monitoring.completed
evaluation.completed
drift.alert.triggered
model.retrained
incident.opened

Inbound events can either carry metrics directly or notify Guardian to retrieve them from a configured source.

Outbound webhook examples

governance.status_changed
consistency.threshold_breached
oversight.review_required
evidence.package.generated
serious_incident.record_opened

Outbound events connect Guardian with alerting systems, workflow engines, GRC platforms, and internal review queues.

Illustrative outbound event

{
  "schemaVersion": "1.0",
  "eventType": "governance.status_changed",
  "tenantId": "example-enterprise",
  "systemId": "wafer-defect-detection-model-01",
  "timestamp": "2026-05-12T10:32:00Z",
  "previousStatus": "GREEN",
  "newStatus": "AMBER",
  "scores": {
    "driftScore": 74,
    "groupConsistencyScore": 83
  },
  "reason": "drift and group-level recall deviation exceeded configured thresholds"
}

What Guardian does not require

A first integration can start from aggregate monitoring signals and model metadata — without changing your existing AI stack.

No replacement of your monitoring stack
No direct access to model weights
No raw training datasets
No raw inference payloads
No intrusive changes to existing ML pipelines
No immediate standardization of all model families under one internal metric framework

Start with one metric source, one integration path, and one governance workflow.

Designed for enterprise integration

Guardian is intended to fit engineering environments where reliability, traceability, and controlled rollout matter.

Authentication: API key, OAuth2, mTLS, or signed webhook payloads
Schema versioning: versioned payloads for safe evolution
Idempotency: duplicate submission protection for retries or repeated batch runs
Retry logic: controlled delivery retries for temporary failures
Observability: ingestion logs, processing status, validation errors, and failed payload tracing
Mapping traceability: every normalized output linked to raw metrics and rule versions
Deployment boundary: support for EU-hosted or controlled deployment models where required
Auditability: threshold rules, mapping logic, and status transitions should be versioned and reviewable

Guardian provides technical evidence infrastructure. It does not determine legal compliance on its own.

Frequently asked questions

Do we need to replace our current monitoring tools?: No. Guardian is designed to sit on top of the monitoring and evaluation systems you already use. The goal is to add a governance and evidence layer, not to replace your existing stack.
Do we need to send raw model data?: Not for a first integration. A metrics-first approach can start from aggregate monitoring signals, model metadata, evaluation windows, and governance-relevant events.
Can we keep our own internal metric mapping logic?: Yes. Guardian can support both native metric ingestion and pre-translated ingestion, depending on how much control your team wants over the adapter layer.
What kinds of AI systems can Guardian support?: Guardian is designed for heterogeneous AI environments, including classification, regression, forecasting, computer vision, ranking systems, and LLM-based applications.
What is the best way to start?: The most effective starting point is a narrow pilot: one model family, one existing metric source, one integration path, and one governance workflow.
Is Guardian a legal compliance determination tool?: No. Guardian provides technical evidence infrastructure and governance workflows. It supports compliance readiness, but it does not replace legal assessment or formal conformity assessment.

Start with one system first

The best way to evaluate Guardian is through a focused pilot: one model family, one existing metric source, one lightweight integration, and one evidence workflow.