From drift signal to new model in minutes
Define trigger policies in YAML. Inferpathio watches the conditions and initiates training pipelines automatically when they're met — with validation gating before anything reaches production.
Three trigger patterns for three failure modes
Not all models degrade the same way. Different failure modes need different triggers — and Inferpathio supports all three.
Metric threshold
Watch a computed metric — PSI score, KS p-value, custom accuracy proxy. When it crosses a configured threshold, training starts. Best for models with clear drift signals.
Data freshness
Trigger when a specified volume of new labeled data has accumulated since the last training run. Best for supervised models with regular feedback loops.
Schedule
Cron-based trigger for models in environments where drift is predictable or business constraints mandate regular refresh cycles. Composable with drift conditions.
Composable trigger conditions
Combine trigger types with AND/OR logic. Require both drift AND minimum data freshness before kicking off an expensive GPU run. Prevent over-training as thoughtfully as under-training.
model: fraud-v4
trigger:
condition: ALL # AND logic
rules:
- psi > 0.10
- new_labeled_rows > 5000
pipeline: fraud-training-v4
validation:
min_f1: 0.90
promote_on_pass: staging
▶ trigger fired: psi=0.14, rows=7234
▶ training run run-9102 startedAutomate the retraining loop
Set up your first retraining policy in 20 minutes.