Vector 1

Demo · Agentic ML

Agentic ML for marketing: from brief to a deployed, monitored model

An agent takes a holiday discount campaign from a one-paragraph brief to a model in production: it checks the data, builds and tests candidates, stops for your sign-off, launches a small canary and keeps watching. The agent does the work. People hold the gates.

The brief

"Our holiday discount email goes to the 30% of the list our current rule picks. Find a better way to choose who gets it, prove it on last year's test, and launch it safely."

Try breaking it

Synthetic customers with a known answer key, generated in your browser. Nothing is uploaded.

The pipeline

    Your sign-off is needed

    Press Run the agent to start.

    Holdout check: profit per 1,000 customers by share emailed

    Runs after the models are built.

    Estimates come from last year's randomized test: half the list got the email at random, so comparing emailed and not-emailed customers inside each targeted group measures what the email added. The answer key is the true value, known only because this is a simulation. Profit counts the margin on extra orders minus the discount given to everyone who used it, including people who would have bought anyway.

    Canary and monitoring

    Runs after you sign off.

    Each week a quarter of the list is targeted by the new model and a quarter by the current rule. In each, 20% of the chosen customers are held back from the email, so the lift is measured, not assumed. The agent pauses the launch if an input drifts away from what the model was trained on, or if measured profit falls below the range it forecast.

    Agent log

    Waiting for a run.

    What this shows, and what it doesn't

    • It shows the practice: a written plan with guardrails, checks before any modeling, candidates compared on data the models never saw, a gate a person signs, a small measured launch and monitoring with clear stop rules.
    • It shows why each gate exists. Turn off the leakage check and the agent picks a model that looks several times better than it is; the drift monitor still stops it before a single email goes out. Add the competitor sale and the model starts losing money in week 3 while every input looks normal; only measuring outcomes catches it.
    • It doesn't show how any particular agent is built. The modeling here is standard and deliberately simple (logistic regression and a two-model uplift estimate), so the steps and the decisions stay visible.
    • It's a simulation. Real campaigns have messier data, smaller tests and slower feedback, which makes the gates more important, not less.