Beauty & personal-care company · via Accenture · 2024
GenAI Video Analysis for Affiliate Creators
AI Engineer
Client engagement. Names, data, and identifying details are withheld; the architecture is described at a generalized level.
Problem
The brand works with thousands of short-video affiliate creators. Teams had to watch videos, read comments, and cross-check sales exports by hand to understand why some videos sell, then coach creators on scripts and product choice.
Approach
The solution went through three iterations, each shaped by brand-team feedback: a sales dashboard with multi-video Gemini analysis; then scene-level scoring, tier coaching, and script generation; then a full creator profile with audience, comment, and style analysis. Every LLM call is constrained by a JSON response schema so the output renders reliably in the app.
Architecture
Hover a component to trace its connections · drag to pan · use + / − to zoom
- Source
- Processing
- Model
- Agent / optimizer
- Data store
- Output
How It Works
- CollectAffiliate sales exports plus each video’s file, metadata, and comments, along with trending keywords and products from public ad-platform insights.
- PreprocessVideos are cached by content hash; face detection, transcription, and fuzzy matching to the product catalog add context.
- AnalyzeGemini reads the video directly, splits it into hook, key-message, and call-to-action scenes, and scores each 0–5 with concrete suggestions.
- ProfileA style profile (selling style, hook and CTA types, tone, editing), comment sentiment and topics, and inferred audience personas.
- CoachA tier engine compares the creator with the next commission tier and names the metric to improve; the app recommends products and drafts scripts that follow ad-claim rules.
- PresentA multi-page Streamlit app with metrics, charts, and word clouds for the brand team.
What I Built
- A sales dashboard over ~43k affiliate videos with filtering and multi-video Gemini analysis.
- Scene-by-scene scoring with a structured rubric: narrative, hook, key message, CTA, audio-visual fit, persuasion, and trend use.
- A creator tier engine, product recommendations, and compliant script generation using top creators’ videos as references.
- Creator profiles with audience personas, comment analysis, and video-style analysis.
- A virality classifier experiment (XGBoost on transcript, visual, and text features) reaching 0.74 accuracy on held-out videos.
Impact
Cut creator review work from hours to minutes and gave brand teams consistent, explainable feedback to coach creators with.