The BI stack is a folder of HTML files.

The problem

Leadership wanted exec-ready views across marketing, merchandising, and product. Enterprise BI meant licenses, a platform queue, and charts that never matched the deck. Notebooks meant nothing a VP could open on a phone at 7 a.m. And every one of these views ends up screenshotted into a deck or a Slack thread, where a sloppy table reads as a wrong number even when it's right.

The build

• Each dashboard is a Python generator that emits one self-contained HTML file. Data embedded, all interactivity client-side. Publishing is copying a file. No BI licenses, no server.
• A generated hub with a staleness dot per tile, refreshed by 36 scheduled jobs.
• A written design standard. Value and delta columns always adjacent. Rate changes in basis points, colored by direction-of-good. One blue for the focus period, one orange for the baseline, everywhere.
• A footing rule. Money rows must match the attribution source of truth to the dollar. People add reports together; a bridge that reads lower is dead on arrival.
• The flagship: a paid-search funnel decomposition that splits any revenue change into eight additive effects summing exactly to the total, so a spend shift into low-converting tactics cannot masquerade as a broken funnel. Replica below: real interface, synthetic data.

What it showed

The week it paid for itself: brand paid-search clicks fell 25%. The room's reflex was "we lost position to a competitor, raise bids." The auction panel said 96% impression share and 0% lost to budget, in both years. The decline was not buyable. Eligible searches were down and click-through was eroding. A six-figure "allocation loss" was reclassified from a budget lever to a demand signal, and the bid increase never happened.

It polices itself, too. One week a view mixed a complete seven-day total with a six-day slice from a stale cache, and a funnel rate dropped about 1,000 basis points across all four tactics at once. Real click-quality shifts are tactic-specific, so that pattern is now written doctrine: denominator artifact, not a real move. When a dead tag made add-to-cart read 726%, the fix was a generic impossibility gate (more adders than viewers) that blanks the metric, shades the band, and self-heals when the backfill lands.

What changed

• The hub is the operating cadence. Stakeholders open dashboards instead of requesting pulls.
• Dashboards get retired when a better cut supersedes them. Curated, not accreted.
• The Wednesday deck runs as one command: detect the fiscal week, run eleven modules against the mart, screenshot each view into deck-ready images. A mid-week correction is a re-run, not a rebuild.
• A retail leader's question from a business review got its own permanently maintained tab.