Challenge
An FMCG brand ran a user-generated content campaign inviting consumers to submit creative entries in audio and video format — a deliberate choice to drive deeper engagement than photo or text submissions. The campaign generated thousands of media entries within the first weeks, each requiring validation against campaign eligibility rules and qualitative scoring against creative criteria for winner selection. Manual review at that volume was infeasible: a panel of judges reviewing each submission would take weeks and introduce significant scoring subjectivity. Existing automation tooling could handle text and image entries but had no meaningful capability to evaluate audio and video content at the semantic and creative level required.
Approach
We built a multimodal validation and scoring pipeline purpose-built for audio and video entry processing. Submitted media files are ingested and routed through a preprocessing layer that normalizes formats, extracts audio tracks from video where needed, and prepares the content for model inference. Each entry is passed to Google's Gemini multimodal model, which evaluates the content against a structured rubric: eligibility criteria (duration, language, product relevance, compliance with campaign rules) and creative scoring dimensions (originality, clarity of message, brand alignment, production quality). The model produces a per-dimension score for each entry along with a structured justification, enabling the scoring to be audited rather than treated as a black box. Entries failing eligibility checks are routed to a disqualification queue with the specific rule violation identified; eligible entries are ranked by composite score to support finalist selection. A configurable scoring weight system allowed the brand's team to tune the relative importance of each dimension without engineering changes.
Outcome
The pipeline validated and scored over 8,000 audio and video entries across the campaign window, a volume that would have required an estimated 1,600 hours of human review at 12 minutes per entry. End-to-end processing time per entry averaged under 45 seconds from submission to scored result. Score consistency — measured by re-running the pipeline on a 10% sample — exceeded 97%, compared to an inter-rater reliability of roughly 68% observed in a parallel manual scoring exercise conducted for benchmark purposes. The brand selected campaign winners from the ranked output with confidence, and post-campaign participant surveys showed no reduction in perceived fairness compared to previous manually judged campaigns.