Ingest. Utility interval data (15-min or hourly), plus whatever connected devices export. Each source arrives with gaps, DST artifacts, and vendor quirks. Step one is normalization to a clean UTC series with gap flags — never interpolating silently, because gaps themselves carry signal (meter swaps, outages, empty house).

Baselining. You need the counterfactual: what would this home have used without the behavior you're evaluating? We blend a weather-normalized regression core (HDD/CDD features) with per-home percentile bands. The model must know when it's uncertain — a tip built on a bad baseline destroys trust permanently.

Anomaly detection. Baseline creep, phantom loads, and schedule mismatches are all deviations from the home's own pattern, not from a population average. Per-home sequence models catch what population statistics smooth over. The output isn't "anomaly score 0.83" — it's "your always-on floor rose 60W in March, roughly $2.50/month."

The explanation layer. Every detection must serialize to one sentence a non-technical person can act on, with a dollar figure and a confidence note. Detection without a decision-ready explanation is a debug log, not a product.

The loop. After a user acts, we track the meter response and score the tip as verified-savings or noise. That feedback is the only thing that makes the next tip better.

EnergyIQ runs this pipeline for real households; the boring-sounding parts (gap handling, uncertainty quantification) are the ones that made it work.