ML Models
ML Models
Manually trigger per-user ML model training. Both models also run on a scheduled DBOS workflow; this is for ad-hoc retraining (e.g. to verify a fix, regenerate a model after data backfill, or test for a single user).
Conflict Model
Pairwise priority model trained from observed time changes: events the user moved AWAY from become winners, events they collided INTO become losers. Used by the scheduling agent to rank candidate slots.
Slot Model
Per-user preferred-slot model trained from the same observed time changes. Predicts which time-of-day buckets a user actually accepts vs reschedules. Hard gate is ≥100 samples (ML_TRAINING_MIN_SAMPLES); ~1000+ recommended for stable models.