All projects
CO
AI & ML2025

Campaign Optimizer

Predict who responds to a campaign, then spend where the next dollar earns most.

Built for: Independent build · public dataset

loading demo…

Runs in your browser with an illustrative logistic model whose weights mirror the direction and rough size of the trained model's drivers. It is a teaching model, not the exported artifact. Move the sliders, then allocate a budget across segments.

The problem

Most campaign budgets are split by habit. The question a model can actually answer is: given what we know about each customer, where does one more dollar of contact cost produce the most responders?

How the real system works

  1. 01Features are built from demographics, tenure, recency, spend by category and past campaign responses.
  2. 02A classifier predicts response probability per customer; the Streamlit app exposes the score and the drivers.
  3. 03Customers are bucketed into segments, and budget is allocated by expected responders per dollar, not by segment size.
  4. 04The CLI exports the trained artifacts so the scoring can run offline in a pipeline.

What you take away

How a propensity score is built from a few features, and why allocating by marginal return beats allocating by headcount.