ML System

Training controls, experiment tracking, feature dictionary, performance monitoring

Intelligence Suite Status

Models Trained

Oldest Model

Max Features

Suite Health

Metric glossary (hover any dotted term anywhere on this page for a full explanation) R² How much revenue variation the model explains (1.0 = perfect) AUC How well the model separates yes/no outcomes (1.0 = perfect) F1 Balance between catching positives and avoiding false alarms WMAPE Average prediction error, weighted by account size MAE Typical dollar error per prediction Silhouette How well-separated the customer groups are (1.0 = perfect) Lift How much better the model's top picks are vs random Checkpoint The best-ever version, kept under lock — production serves the champion Vintage replay Retrain in the past, grade on the known future

Model Suite

What this is

Twelve models that analyze the business from different angles — account revenue, churn, segments, expansion, event and subscription renewals, individual deals, individual contacts, and never-bought prospects. Together they produce 15+ intelligence dimensions per account so the business can act on data instead of intuition. Every model is graded against an honest baseline (naive carry-forward, base rate, or the CRM's own numbers) in the vintage replay, and a monthly retrain job keeps weights fresh behind the checkpoint gate.

How they're organized

The suite trains in a cascade. Segmentation (M3) runs first and assigns every account to a behavioral cluster per sales team. Those segment labels then flow as input features into M1, M2, M4, M7, M8, and M14, grounding their predictions in the behavioral patterns M3 discovered. M6 is independent (procurement data), M10–M12 score pairs and deals rather than whole accounts, and M13 scores individual contacts behind a serving gate.

M3 Segments → M1 Revenue M2 Retention M4 Expansion M7 Portfolio M8 Cross-Sell M14 First Buy | M6 Markets M10/M11 Renewals M12 Deals M13 Contacts

What each model does

M1 — Revenue Forecast: Forecast = last year's revenue + a gated adjustment (anchored delta), split into quarters by sub-models. Drives revenue-at-risk and the quarterly outlook.
M2 — Account Retention Risk: Scores the probability an active account churns to zero revenue. Combined with M1, produces dollar-weighted risk.
M3 — Account Segmentation: Clusters accounts into behavioral segments (per sales team) using 250 features. Foundation for downstream models.
M4 — Category Expansion: Predicts single-category accounts that will adopt a second product category. Identifies upsell readiness.
M6 — Event Market Intelligence: Rules-based engine that scores state×category procurement momentum and aligns sponsors to market tailwinds. No ML — pure analytics.
M7 — Event Portfolio Expansion: Predicts which event sponsors will diversify from one event type to many. Growth signal for the events team.
M8 — Cross-Sell Graduation: Predicts single-line accounts that will buy across multiple product lines, and recommends the most likely next category.
M10 — Event Series Renewal: Per (sponsor × event series): will they re-sign next year? Feeds Event Resigns' "Resign odds".
M11 — Subscription Renewal: Per (account × subscription family): renewal odds from usage telemetry, with gone-dark accounts capped.
M12 — Opportunity Win Probability: Per open deal: win odds that beat the CRM's stage percentages head-to-head. Deal × odds = expected value.
M13 — Contact-Product Propensity: Ranks the people inside an account per product family. Serves only while it beats the deterministic ranking (gate).
M14 — First Sponsorship: The prospecting model: never-sponsor companies scored for their first event buy. Rides Atomic Inventory.

Reading the cards

Each card below shows a model's current state. The headline is the key metric at a glance — R² for regression, AUC for classifiers, Silhouette for clustering. Metric chips (colored badges) break this out: green is strong, amber is acceptable, red needs attention. The narrative block explains what the model does, how it's performing, and what it reveals. Feature importances show which input signals the model relies on most (for M3 this shows per-team cluster quality; for M6, market momentum and alignment distributions). Use Configure to tune hyperparameters, History to see how performance changed across runs, Log for raw training output, and Debrief to get an LLM-generated analysis with actionable config recommendations.