Product
From drift signal to resolved.
Inferpathio monitors every production model's prediction distribution. When PSI, KL divergence, or Wasserstein distance crosses your threshold, a governance event fires — entering a policy-defined approval queue, triggering your training pipeline on approval, and recording every step as an immutable audit event. Nothing in this pipeline is manual. Nothing is undocumented.
The Full Lifecycle
Every stage, governed and logged
Inferpathio covers the entire model lifecycle from initial drift signal to post-deployment audit review.
Drift Detection
The Python SDK logs prediction batches continuously. When Population Stability Index (PSI) exceeds your configured threshold, an automated governance event is created. Supports PSI, KL divergence, and Wasserstein distance metrics.
Approval Queue
The drift event enters the retrain approval queue. Approvers are notified via Slack, PagerDuty, or email based on your RBAC policy. SLA timers begin. If approval expires, the event escalates automatically.
Policy Evaluation
Your YAML policy is evaluated against the event. Threshold breaches, accuracy floors, and approval roles are all checked before a retrain pipeline is permitted to trigger. Policy version is recorded immutably.
Automated Retraining
On approval, a signed webhook payload fires your training pipeline in Airflow, Prefect, GitHub Actions, or any HTTP endpoint. The pipeline run ID is tracked in Inferpathio's audit log from trigger to completion.
Deployment & Promotion
The retrained model candidate enters a promotion workflow. Shadow mode evaluation, canary splits, and full swap patterns are all supported. Inferpathio tracks the version tree from experiment through every promotion stage.
Rollback Authority
If post-deployment accuracy drops by your configured threshold within the rollback window, an auto-rollback triggers. Manual rollback to any ancestor version is available in under 60 seconds with full lineage preserved.
Policy Engine
Governance as code
Every retrain decision is governed by a versioned YAML policy file stored in your version control system. Policies define threshold triggers, approval chain requirements, auto-rollback rules, and SLA timeouts.
- Per-model policies — different thresholds for credit models vs recommendation engines
- Multi-metric triggers: PSI, accuracy floor, prediction volume drop
- Auto-rollback rules with configurable accuracy drop thresholds
- Policy version stored in every audit event — complete traceability
# Full policy spec for credit-risk model
model: credit-risk-v3
environment: production
retrain_trigger:
psi_threshold: 0.08
accuracy_floor: 0.91
min_prediction_volume: 5000
evaluation_window_hours: 24
approval:
required: true
approvers:
- role: ml-lead
min_approvals: 1
- role: compliance-owner
min_approvals: 1
timeout_hours: 24
escalation_role: platform-lead
pipeline:
webhook_url: "${RETRAIN_WEBHOOK_URL}"
timeout_hours: 6
auto_rollback:
on_accuracy_drop: 0.03
within_hours: 1
notify:
- slack:#ml-alerts
- pagerduty:ml-oncallIntegrations
Governance that sits on top of your existing stack
Inferpathio is not a replacement for MLflow, SageMaker, or Airflow. It is the governance layer above them — reading artifacts, writing audit events, and routing approvals without requiring you to change your training pipelines.
Reads artifact URIs, run IDs, and model registry transitions. Governance attaches to your existing MLflow runs — no schema changes required.
Connects to SageMaker inference endpoints via CloudWatch metric streams. Captures prediction volume, latency drift, and feature distribution shift without SDK code in your serving container.
Monitors HuggingFace Inference Endpoints. Tracks model version changes and routes deployment events through your governance pipeline.
Sends a signed webhook to your DAG trigger endpoint on retrain approval. The Airflow run ID is logged back into the Inferpathio audit trail via a sensor task.
Uses the repository_dispatch event to trigger training workflows. Approval events include the policy version and approver identity as input parameters.
Native Prefect block and deployment trigger. Retrain flows receive governance metadata as Prefect run parameters; flow run IDs are tracked in the audit log.
Configurable per-model channel routing. Approval requests arrive as interactive messages — approvers can respond directly in Slack without opening the dashboard.
Routes critical drift alerts and auto-rollback events to on-call rotations. Different PagerDuty service keys per model environment (staging vs production).
Connects to MLflow on Databricks via the MLflow Tracking Server API. Works with Unity Catalog model registry for artifact lineage.
Governs Vertex AI model endpoints and pipeline runs on GCP. Reads endpoint prediction logs via Pub/Sub; writes governance events back to Vertex Metadata.
Okta, Azure AD, and Google Workspace identity providers supported. Role assignments can be synced from IdP group membership via SCIM provisioning.
Any HTTP endpoint can receive governance events. Payloads are HMAC-signed with a per-integration secret; retry logic with exponential backoff is built in.
Walk us through your stack, we'll show you where governance fits
Demos are technical — bring the person who manages your MLflow setup or SageMaker endpoints. We'll show you how the SDK attaches, what a policy file looks like for your model type, and what the audit export produces.