Smartex — Creative Intelligence Copilot

HackUPC / Smadex challenge. Explains why some mobile ad creatives perform better than others, detects fatigue, and recommends next actions — powered by a LightGBM v4 scorer and SHAP explainability pipeline.

Quick start

Frontend (React + Vite)

cd src/frontend
npm install        # first time only
npm run dev        # starts at http://localhost:5173

Backend (FastAPI)

uv sync                                              # install deps (first time)
uv run uvicorn src.backend.main:app --reload         # starts at http://localhost:8000

API docs available at http://localhost:8000/docs once running.

Notebooks (Python analysis + modeling)

uv sync            # install Python deps
jupyter lab        # open notebooks/

Notebooks run in order:

• 01_eda.ipynb — exploratory analysis
• 02_model_scorer.ipynb — LightGBM v1/v2
• 04_multimodal_features.ipynb — v3/v4 scorer + SHAP + advantage function

Screens

Route	Screen
/	Advertiser Selector
#/advertiser/:id	Portfolio Overview
#/campaigns/:id	Campaign View
#/campaigns/:id/fatigue	Fatigue Detector
#/campaigns/:id/recommendations	Recommendations
#/advertiser/:id/reviewer	Creative Reviewer
#/advertiser/:id/chat	Copilot Chat

Stack

• Frontend: React 18, Vite 5, Plotly.js (basic bundle)
• Modeling: LightGBM, SHAP, scikit-learn
• Data: 1,080 synthetic creatives across 36 advertisers and 180 campaigns