Nothing in this post has been built. It’s a plan, written out so you can check my thinking.
Where this came from
The team at a frozen-food brand wants to ask plain-English questions about how their Walmart business is doing. The answers live in SPINS syndicated data. Their in-house IT doesn’t know how to build it. We’ve talked it through. Nothing has been built.
How I’d build it
- A scheduled SPINS export. The data starts as a file out of SPINS Satori.
- A semantic layer. Dollars, units, velocity, distribution and share, over 4, 12 and 52 weeks, against a year ago. The model never has to guess what velocity means for this brand.
- Text to SQL, with the numbers shown. A question becomes a query. The answer comes back with the figures it used, so anyone can check it.
- A simple chat screen behind the company login.
Why show the numbers
A chat that answers in sentences is easy to trust and hard to check. If the answer comes with the query and the figures behind it, a person on the team can spot a wrong number before it goes in a deck.
The caveat
The software is the fast part. The slow part is data access and licensing with SPINS. Before anyone builds anything, the brand needs to know what its SPINS agreement lets it pull, whether the retailer in question is in that data at all, and where the data can live.
I’d rather say that up front than find it out in week three.