The thing about working with data scientists is that the hardest part of the engagement has nothing to do with the data. A model is only as useful as the question it's built to answer. And writing a question that a model can actually answer — specific enough to be solvable, broad enough to matter — is where most technical collaborations quietly fall apart.

At Walmart, I was brought in as Technical Program Manager on two parallel data science workstreams as part of the Live Better Hackathon — both emerging from the same underlying pressure: Walmart had made ambitious, public sustainability commitments, and needed to understand whether the operational data it was already sitting on could make them real.

My role across both workstreams was the same: partner with business stakeholders to identify and sharpen the problem statement, develop product requirements the data science team could execute against, prep and validate data sets, and manage the technical sprint from framing to model to output. The design thinking wasn't in the interface. It was in the question.

Walmart's Quality Control managers had accumulated years of anecdotal evidence suggesting a pattern: weather events at growing locations — heavy rainfall, temperature extremes, high wind velocity — seemed to correlate with produce rejections downstream at distribution centers. Rejections meaning: product that travels the full length of the supply chain only to be deemed unacceptable upon arrival, often ending up as waste.

The hypothesis wasn't controversial. The problem was that it had never been systematically tested. Anecdote and data are different things, and only one of them can drive a supply chain decision.

We built the problem statement around a specific question: could we predict inbound produce rejections — before product ever reached a distribution center — based on weather conditions at the source fields? If yes, smarter sourcing decisions upstream could reduce rejections, improve instock rates, and divert product from landfill before it became waste.

The data architecture combined two sources: Walmart's internal QC inspection records — including inspection date, DC number, average core temperature at inspection, supplier name, origin country, and rejection result — with NOAA's publicly available weather data, pulling temperature, precipitation, and wind speed for specific growing regions. We selected four to six commodity and growing location combinations based on weather sensitivity and rejection history, deliberately mixing domestic and international origins to give the model breadth.

The sprint produced a predictive model capable of flagging spoilage risk pre-shelf — not just a binary up-or-down signal, but a graduated risk estimate that could inform sourcing volume and routing decisions in advance of a product ever leaving the field. The model gave Walmart something its QC managers had never had: foresight with evidence behind it.