Predicting ad performance before the money is spent

The problem

Advertising spend is committed before anyone knows whether a creative works. The feedback loop is expensive and slow: run the ad, wait for the numbers, learn something after the budget is gone.

Sporelogic wanted to move that judgment earlier, to say something useful about a video advertisement’s likely performance from the asset itself.

The approach

I worked on this project during my internship. The core idea was that a video ad carries most of its persuasive signal in two places: how it sounds and what it says.

That meant treating it as a multi-modal problem from the start, pulling acoustic features from the audio track (pace, energy, tonal variation) and linguistic features from the transcript (structure, framing, where the call to action lands), then modeling performance against both.

The interesting engineering was not the model. It was the pipeline: taking inconsistent real-world video files and turning them into a feature set clean enough that the model was learning about advertising rather than about encoding artifacts.

What I took from it

Most of the value in an applied ML project sits in the unglamorous half: data quality, feature design, and being honest about what the model does not know. It is the same lesson I bring to client work. The model is rarely the hard part.