Project 03 · Case study
AI Video Ad Concept Analyzer
GPT-5-powered evaluator that scores ad concepts across six dimensions. Built to test whether LLM feedback on creative work can be useful or just drifts with temperature.
Case study
The problem
Evaluating video ad concepts is mostly vibes. I wanted to see whether you could get a model to give usefully consistent feedback across six axes, or whether the output would just drift with temperature.
The approach
GPT-5 under a carefully structured prompt that returned JSON across six dimensions (creativity, emotional impact, brand alignment, clarity, memorability, target relevance). CLI and FastAPI interfaces. Retry logic and schema validation on the output, because JSON-mode alone doesn't save you from a malformed response.
What worked
Output was consistent enough across reruns that the scores were actually useful as a second opinion, not random noise. The FastAPI layer made it drop-in for an existing creative review workflow.
What I'd do differently
Six dimensions was too many. Three would have been sharper. The model hedges when you ask it too many things at once.
More detail
A structured-prompt system that evaluates video ad concepts across creativity, emotional impact, brand alignment, clarity, memorability, and target relevance. CLI and FastAPI interfaces, with schema validation and retry logic on the model output.