Shipping ML that marketers actually use

Appvertiser had the data, the engineers, and the ambition. What they needed was a prediction system that didn't just output numbers — but output numbers that the marketing team would actually act on. That gap, between a working model and a trusted recommendation, is where most ML projects die.

The model is 20% of the work

The first model we built performed well on the validation set. The marketing team looked at it, nodded politely, and kept doing what they were doing. The predictions were good, but they weren't trustworthy. No one could answer the question 'why did the model say that?' — and no one bets a campaign budget on a black box.

Trust comes from transparency, not accuracy

Every recommendation now ships with: the top three factors driving the prediction, how similar campaigns performed historically, and a confidence band that's explicitly communicated. The raw model output is available for power users, but the default experience leads with reasoning.

• Recommendation cards show contributing factors in plain language
• Historical parallels link to actual campaign data, not summaries
• Confidence bands display as ranges, not single numbers
• Every prediction can be traced to its training data lineage

The feedback loop is the product

The system improved because the marketing team could mark predictions as 'useful,' 'wrong,' or 'applied' — and those labels fed directly into model retraining. After three months of this loop, the predictions weren't just more accurate; they were dramatically more aligned with how the team actually thought about campaigns. The model didn't train the users. The users trained the model.