One project used anonymized B2B manufacturer data for revenue forecasting, customer scoring, and prospect scoring. The other uses years of my own paid-ad data across paving, garage door, and Heaviside-managed accounts.
The useful part is not the model name. It is whether the model changes what someone does next.
This page is about source data, baselines, validation, and whether the work made a real decision easier.
Here's a summary of this page by my AI Avatar
Start with the underlying records, not screenshots or stale dashboard summaries.
Measure against a simple baseline before claiming a model improved anything.
Use holdouts, walk-forward checks, and segment reads where the data calls for it.
The model matters only if it changes a decision, alert, report, or review loop.
Multi-year revenue forecasting program for an anonymized B2B manufacturer. XGBoost on 27 engineered features, walk-forward validation, state disaggregation that cut MAPE in half, autoresearch agent campaign that drove global mean MAPE to 3.77%.
Customer segmentation, prospect lookalike scoring, repurchase prediction, and customer-potential scoring across ~34K customers and a six-figure external-institution reference universe. Surfaced an eight-figure identified growth opportunity across the top-scoring accounts.
What it actually takes to run a real ML system in production. MLflow registry, FastAPI service layer, React frontend, Docker + Coolify + Cloudflare, JWT + SES email auth, quarterly retrain runbook, drift monitoring, autoresearch off-net runtime.
Years of paid-ad data across 165+ accounts in paving, garage door, and Heaviside-managed verticals. Warehoused, vertical-labeled at ingest, ML feature view per account-day, production scoring engine that writes risk scores and threshold alerts on every warehouse run. Same methodology as the anonymized B2B forecasting work, applied to my own ad data.