User funnel& upgradepredictor
A leakage-safe upgrade-prediction model and a strict-nested 6-stage funnel, built end-to-end in Zerve. Drag the 3D manifold, score any user, retune the funnel rules — every prediction is explainable.
The canvas, in your browser
Every block here is a real Zerve canvas node. Coordinates, edges, descriptions all read straight from canvas.yaml. Click a block to pull its live matplotlib figure or variable from the deployed FastAPI — same calibrated XGB ensemble, same cohort tables, same K2 strategies that the data scientist sees inside the canvas.
Live inference · v3 ensemble
Pick any test-set row index — the request flies through beta-zerve.hub.zerve.cloud, which calls zerve.variable("Train Model v3", "models") on the canvas and runs predict_proba across all 3 calibration folds. The probability you see is generated server-side, not baked into a static JSON.
Hit the v3 ensemble in real time
Each click goes to the deployed FastAPI on beta-zerve.hub.zerve.cloud, which calls zerve.variable("Train Model v3", ...) on the live canvas — same calibrated XGB ensemble you see in the DAG above.
K2-Think strategies · per-segment
14 v4 funnel segments. Each one shipped through the Build Strategies node, which prompts K2-Think with the segment's behavioral profile and a Zerve playbook excerpt. The 3 ranked actions, target filters, expected uplift, and ROI multipliers come back in real time from the deployed canvas.
Insights card · the fan-in
The final canvas node — fan-in from Diagnose v3, SHAP v3, Compare Models, Per-Segment Performance, Build Strategies, and ROI Ranking. The PNG and the text both come from the same canvas variables, fetched live.