Overview › Framework
Framework
The map — every step from raw data to deploy, as a shared platform (areas 1–4) plus a per-model spine (areas 5–6). Every box links to the page that owns it. The rules behind each step live in the Methodology.
Cross-model platformBuilt once · every model reuses it · Roofing · Apollo · Olivia · Garage · Windows
1Data foundation
1.1
Scope & data audit gate
Profile every source — funnel (leads · appointments · deals) · fulfillment · Silver_FA · Gold_Permits · deeds · external — quantify join keys, date grain and deal volume, then freeze the label spec. Nothing downstream starts until this one-time gate passes. The corrupt
event_date is the #1 label-timing risk.1.2
Data extraction
Where each dataset comes from — the raw vendor feeds and reference dictionaries.
1.3
ETL & quality
Raw vendor data → clean Gold. How the engine transforms permits and parcels, and what every Gold field means.
2Target definition
2.1
Labeling
Define valid positives — shared methodology, model-specific label. Roofing: which permits count as a roof event. Apollo: any property sale (y_sold). Olivia: a dealable transaction via funnel decomposition — the client's closed deals are positives, every other transaction is unlabeled (positive-unlabeled), corrected with a true prior.
2.2
Coverage
Where "no permit" really means "no roof" — the join between permits and parcels, and where the vendor has visibility. Bounds where a negative label is trustworthy.
3Feature engineering
3.1
Enrichment
As-of-T0 property state — every value computable from the snapshot alone, leakage-clean.
3.2
Feature families
Internal × external signal families — the shared feature pool; each model picks its own driver clusters.
4Assurancecross-cutting
4.1
Audits & validation
The continuous, independent check on every number — standing audits, the QA battery and the findings log. Spans the whole build (distinct from the one-time data-audit gate in 1.1).
Per-model spineEach model defines & runs its own, on top of the platform above
5Modeling
5.1
Folds
Walk-forward folds + case-control sampling — no training on future data.
5.2
Training
Gradient-boosted machine — a single global model plus a true-prior correction. Olivia trains the two funnel stages here.
5.3
Calibration
Turn a ranker into a forecaster — map scores to honest 30/60/90-day probabilities.
6Delivery
6.1
Output gate
Recent-roof exclusion, dedup and a health check before anything ships.
6.2
Cadence & delivery Olivia
Turn calibrated scores into each client's monthly list — a policy layer, not a new model. Four levers: which N to send (variable capacity) · burnout / touch-ledger suppression · timing / ripeness · per-client buy-box fit.
6.3
Deploy & monitor
Backtest and value metric — for Olivia a standing value-audit harness: does the as-of-T0 ranked list recall the deals that actually happened (external included), vs Alpha and the market base rate? Plus retrain cadence and the live coverage surface.
The 8020IQ model framework · mirrors notes/Roofing/PROGRESS_NOTEBOOK.html · roofing = worked example; every model follows the same areas (cross-model platform + per-model spine). Models live on the Models page.