UMDB — Monetizing the Workflow, Not the Network

Context

As Data Product Analyst I built the analytics layer underneath Carfax's product organization: ETL pipelines in Azure Synapse that turned raw data into trusted, analysis-ready datasets, running across 35B+ records from 260K+ sources. The visible output was four product dashboards — vehicle-valuation usage, browser-extension analytics, consumer analytics, and iOS value metrics.

The problem

Feature-level usage was invisible. Everyone knew the product was used; nobody could see which features actually carried it. So roadmap arguments were won by whoever argued hardest, and there was no evidence to justify killing a feature or doubling down on one. Prioritization was running on conviction, not data.

What I did

Built ETL pipelines in Azure Synapse producing analysis-ready datasets the whole product org could work from, instead of one-off extracts.
Ran large-scale analytics across 35B+ records and 260K+ sources to surface genuine feature-level utilization, not vanity totals.
Used cohort analysis to connect usage to retention — tying what people did to whether they stayed, rather than just counting clicks.
Shipped self-serve dashboards that retired the ad-hoc request queue, so product managers could answer their own questions without waiting on me.

The decision I’d defend

The finding was blunt: a small handful of features carried the majority of active usage, and much of the surface was dead weight. That forced a choice — invest in the few features that worked, or spend the same effort trying to fix the ones users had already ignored.

Rejected: spreading engineering across the underused surface to "rescue" it.

I argued to double down on the proven features. Users had already voted with their time; concentrating engineering where that signal was strongest returns faster than betting on features they'd walked past once already. Reviving a dead feature is a hypothesis — reinforcing a live one is a near-certainty.

Outcome

The analysis moved prioritization from opinion to evidence and fed directly into roadmap decisions alongside the product managers — the bridge from measuring the product to helping shape what it does next.

What I'd do differently

I'd push to instrument features at ship time rather than reconstructing usage after the fact. Retrofitting analytics onto features that launched blind cost weeks a simple logging spec up front would have saved.

Product specifics and outcome metrics are covered by confidentiality. Happy to walk through the details in conversation.