Messy problems in. Clear decisions out.

Dealers pay to join the dealer program, and account managers need to show them what they get back. That answer was spread across a dozen sources, and the monthly update took days of manual work with close attention to detail.

I built the first scrappy version in 2018 with Excel, Alteryx and Tableau. In 2024 I took it back and rebuilt it: 4 internal cloud sources plus 8–10 partner feeds, cleaned and joined into one monthly dataset on program use, purchases and ROI for every dealer. I also set up automated national-comparison reports for independent distributors who don't have access to our internal dashboards.

A 3,000–5,000-dealer network in one interactive dashboard. Account, territory and regional managers use it to run their regions and in one-on-one reviews to renew dealers and show them where to improve. The monthly refresh went from days to one morning.

The old version was a separate table for every source, rebuilt each year, so it could only ever show one year. I restructured everything around a single dealer source of truth and a date key. Every source now lands in one table, and each month's data is appended instead of rebuilt, so the dashboard runs year over year. It works at the month level; managers asked for year-to-date, so that's what it shows, and switching back is easy.

Twenty-plus recurring data flows fed the team's Tableau reports. Many still needed someone to pull exports, rerun steps by hand and check the results every cycle, and fragile workflows broke often.

Led my team's move from Alteryx to Dataiku for scheduled, more stable pipelines. As part of Trane's Dataiku Champions group I tested new features before company rollout, taught the team the tool and was their go-to contact while they learned. For the heaviest monthly projects I built end-to-end Power Automate → Dataiku processes, so files are collected, prepped and published without anyone touching them.

Dozens of team hours saved each month on routine updates, about 5 hours a week for me, and some monthly projects went from multiple days of manual work to hands-off runs.

Campaign data lived in separate platforms. GA4, Campaign Manager, Google Ads, Meta, LinkedIn and BrightEdge each told part of the story, and none of them connected to leads or dealers.

Designed the data structures that join those sources into one clean model. Brought in new sources like Google Ads and BrightEdge, automated the prep in Dataiku, then built performance dashboards and trained marketers to use them on their own.

Marketers measure outcomes by tactic and channel without waiting on an analyst, and leadership sees one version of return on marketing spend.

Homeowners use the dealer locator to find someone to call. Ranking by distance alone didn't surface the best dealers, and the data suggested lots of frustrated, repeated searching.

Designed a dealer score from customer feedback, response speed, Google reviews and dealer level, so the best-fit dealer shows first in each ZIP code. Then I dug into how people actually use the tool and found the site was logging a search on every page load, so thousands of "frustrated searches" weren't real.

Simpler ranking logic and accurate reporting. It also set the locator strategy: always show three choices, and when no dealer is available, a general contact card takes the empty spot. That was tested and is live today.

To connect website visits to dealer phone calls, every session gets its own tracking number. We moved from a vendor pool of millions of numbers to our own Twilio setup with only a few thousand that rotate. Call volume dropped hard, and the question was whether we'd lost real customers.

The rotation logic was built by developers; I owned the analysis. I compared call patterns, durations and completion before and after the switch to find out who the "lost" callers actually were.

The lost calls were almost all bots, random dials and people who weren't looking for a dealer. Hitting a number from a small pool by chance is far less likely. Dealer call completion improved and average call time went from under 30 seconds to about a minute. It changed how the business defines a healthy call volume.

Most homeowners have no idea what size system they need, and that uncertainty stalls them before they ever call a dealer. The official load calculations need details a shopper doesn't have.

Researched how dealers size systems, then pushed to get real-world housing listing data with known system types. I used machine learning to weight each factor (square footage, ceilings, insulation, windows, orientation, occupants, age) into a points system. Homeowners answer in simple ranges instead of exact numbers. The output was good, better and best options (14, 17 and 20 SEER) with a cost-to-own chart using local weather and energy prices, plus a call-a-dealer button.

The proof of concept became the base developers built the website tool from. I've rebuilt my version as a working calculator on this site.

We ran dozens of tests a year, but campaign traffic was thin, so down-funnel results were sparse and often inconclusive. One test removed a "contact dealer now" card from product category pages. It showed no major impact, so the card came out.

About a month later, leads from that area were way down. I went through the lead data and found that the source had dropped out as expected, but those calls never showed up anywhere else. People weren't finding another path. They were just leaving.

The card went back and lead volume recovered. The lesson I carry: an inconclusive test isn't a neutral one, so watch the downstream numbers after anything ships.