Contact us
A U.S. marketing intelligence agency measures TV, radio, digital and weather-driven demand for advertisers, and every client's data arrived with different column names, on a slow front end with charts clients couldn't question. We rebuilt the front end, used Claude 3.5 Sonnet to map each client's fields into one schema while the attribution stays in code, and let clients ask their charts questions.
The data arrived in a different shape every time.
Clients couldn't interrogate the answer.
As the agency reported.
The agency helps advertisers see what their spending actually does across television, radio, digital channels and weather-sensitive demand, using attribution and demand forecasting. Its clients sent data with column names that varied widely, so every dataset needed manual corrections before anything downstream could run, and errors and delays crept in.
The front end was slow, and its charts offered little to interact with. Clients could look at a chart, but couldn't ask it why, compare two views side by side or get a summary, so decisions and reports took longer than they should.
The design drew one line through the work. Spotting that two differently named columns mean the same thing is what a model does well, so the model maps the fields. The attribution calculations are business logic, and they stay in code, where they can be tested and trusted.
3 decisions, in the order we made them. Each one says what we did and why, with what it did drawn beside it.
Why: when a model decides what a column means, someone can check it in seconds. Nobody can check a model's view of what a TV spot was worth.
“Armakuni goes above and beyond to deliver and copes with mistakes instead of hiding them.”
As the agency reported.
The weather module for the remaining clients, and field mapping extended to more client datasets once each is validated.
This project came before our current offers. Here's how we'd approach the same work now.
Column mapping is a narrow, repetitive task, so AI Inference Engineering would give it a model the agency owns on Amazon SageMaker AI, at about a tenth of the price, with the hard cases still going to a frontier model through Amazon Bedrock. On AK AI Factory, every mapping someone already corrected by hand would become the golden set each change must pass, plus a sealed set nobody tunes to, and Axiom would keep one catalogue on Amazon DataZone, so every AI agent and every chart reads "lead", "spend" or "lift" the same way.
A free workshop: a few hours with the architects who'd do the work, on your own systems. No cost.
Book a free workshop