Open to interesting problems
Enterprise AI doesn't fail on models.
It fails on the data underneath.
I'm Kevin Richard. Sixteen years in digital analytics, working on what gets captured, whether it's correct, and whether anyone can make a decision from it.
Approach
Three layers, in order
Skipping one doesn't save time. It moves the failure somewhere more expensive.
01
Instrumentation
Most data problems are collection problems in disguise. If what's captured at the source is wrong, nothing downstream can be trusted.
02
Pipelines
Moving data is easy. Moving it so failures are visible and yesterday's number still reconciles today is the part teams skip, then pay for.
03
AI readiness
Models amplify whatever they're fed. Defensible output depends less on the model than on lineage, definitions, and governance nobody wants to own.
Projects
Things I've built
Side of the desk, mostly — tools built because the manual version wasn't scaling.
LLM-simulated survey respondents
Adobe Analytics debugger
Instrumentation audit for retail eCommerce
Background
Sixteen years, agency then in-house
Syncapse, Edelman, J. Walter Thompson, Track DDB, Publicis, and now LCBO. Full history and dates live on LinkedIn.
About
Where I come from
Toronto, Ontario
Agency years on accounts where the reporting had to be right — Samsung through global device launches, Air Canada's display program, Volkswagen and Stellantis through site rebuilds. Then in-house at LCBO, where I stopped handing analysis over for someone else to implement.
A dashboard everyone relies on turns out to rest on a tag that stopped firing months ago, and nobody noticed because the number still looked plausible. The fix is almost never downstream.
AI raised the stakes on that. A model trained on unreliable data gives unreliable answers with more confidence, and less visibly, than a spreadsheet ever did.