Problem
An entertainment-industry client produced large volumes of creative assets and promotional media every quarter and needed a way to predict which ones would resonate with their audience before committing distribution and paid-media spend. Existing analytics surfaced post-launch performance but offered no signal during the pre-launch window where decisions actually get made.
Approach
Built a multimodal creative-intelligence model that combined visual and audio embeddings of the asset, structured catalogue metadata (genre, talent, format, release window), and historical audience-performance signals. GPT-4V, Gemini, and Claude were used in a layered pipeline — first to characterise assets along audience-meaningful dimensions, then to project against historical performance distributions for comparable content. Output integrated into the client's existing media-analytics workflow as a pre-launch scoring layer.
Stack
GPT-4V · Gemini · Claude · multimodal embeddings · catalogue metadata pipeline · entertainment-industry creative-evaluation framework
Outcome
Pre-launch creative-performance scoring became a structured input to the client's greenlight and distribution decisions, replacing what had previously been a mix of intuition and post-hoc analytics. Framework portable to adjacent problem classes across media analytics — content commissioning, format selection, thumbnail and key-art testing — wherever historical performance distributions exist as a learning signal.