Tracked KPIs — a gold layer that AI can actually find

[roman1887]

The sidebar entry is "Tracked KPIs" (renamed from Scheduled KPIs in the UI; README/docs still say "Scheduled KPIs" — that's catching up). Naming aside, the substance:

Write SQL once. On a 1–24h schedule, results materialize as a real typed Postgres table in the kpi_data schema. The table is registered with FAISS, so the next time someone asks "what was our daily active users last week" the model finds the materialized table instead of generating a from-scratch query against raw events.

This is the same gold-layer pattern as Databricks gold tables or Power BI import mode, but on plain Postgres + Trino. No proprietary format, no separate compute.

The collected_at column gives you free time-series, and the same KPI can feed a Predictive Intelligence pipeline that auto-retrains every time new data lands. The chain works:

raw data → Tracked KPI (gold) → Prediction pipeline → prediction_data table (also FAISS-indexed)