Portfolio
AI Strategy & Innovation
This is where I keep my work at the intersection of artificial intelligence and communications. I am earning my MS in Artificial Intelligence Strategy and Innovation at Wake Forest University, where I learned to build and evaluate machine learning models and, just as important, to translate what the data finds into decisions a business can actually use. Below is a full project, start to finish.
Graduate Coursework / Wake Forest University
Predicting Organic
Purchases
A machine learning consulting project, from raw data to a plan the business can run.
The Brief
A supermarket needed to spend smarter.
The client runs a loyalty program and wanted to know which customers are actually likely to buy organic products, so promotion budget lands on the right people instead of everyone. My job was to turn their customer data into that answer.
The Approach
Five stages, each one building on the last.
Milestone 01 / The Baseline
Logistic Regression
I started simple on purpose. A reduced model came down to three signals that mattered most: a customer's affluence, age, and gender. Easy to read, easy to explain to a marketing team.
A solid, honest starting point. The model was right when it flagged a buyer about 70% of the time, which set the bar everything else had to beat.
Milestone 02 / Going Deeper
Classification Tree
Next I built a decision tree and tuned it with a grid search so the settings were chosen by the data, not by guesswork. It pushed precision higher, but it came at a cost.
More confident when it spoke, but it stayed quiet too often. That low recall meant it missed a lot of real buyers, which is the kind of trade-off you have to catch before it costs the client.
Milestone 03 / The Real Lesson
When a perfect score is useless.
I built two ensemble models. Random Forest scored a stunning 0.974 precision, almost never wrong. The catch: it only caught about 2 in every 100 actual buyers. Beautiful on paper, worthless in practice.
So I chose Gradient Boosting instead. It traded a little precision for far better balance, and balance is what a real marketing campaign actually needs.
The Decision
Four models, side by side.
| Model | Accuracy | Precision | Recall | AUC |
|---|---|---|---|---|
| Logistic Regression | 0.808 | 0.701 | 0.394 | 0.804 |
| Classification Tree | 0.789 | 0.744 | 0.228 | 0.742 |
| Random Forest | 0.758 | 0.974 | 0.022 | 0.788 |
| Gradient BoostingSelected | 0.812 | 0.732 | 0.380 | 0.809 |
Precision is how often the model is right when it flags a buyer. Recall is how many real buyers it manages to catch. Gradient Boosting won because it was strong on both at once, not just one. Logistic regression figures are in-sample.
Milestone 04 / Segmentation
From numbers to people.
Prediction tells you who. Segmentation tells you who they are. I used k-means clustering to sort all 22,223 customers into four personas a marketing team can actually speak to.
Milestone 05 / Deployment
A model only matters if it runs.
The last stage was an MLOps plan: how this model leaves the notebook and lives inside the business. Most machine learning projects fail here, not in the math, so this is where strategy earns its keep.
What This Taught Me
Learning to speak data without losing the story.
I came into this program as a communications strategist, not a statistician. What I took from this project was the part that sits between the two: reading what a model is really telling you, knowing when a great-looking number is a trap, and translating all of it into a decision someone can act on.
Thank You
Molly McWilliams
Wilkins
Be the Light. Walk in Truth. Grits and Grace.