Tigri

Inspiration

Creative fatigue is expensive and usually detected too late. We wanted a tool that gives non-technical advertisers quick, practical answers: what is working, what is fading, and what to do next.

What it does

Smadex Creative Intelligence is a 3-level dashboard powered by two ML models.

Advertiser overview

Benchmark KPIs (CTR, CVR, ROAS) vs peers in the same vertical
Peer ranking chart
Country performance map
Clickable campaign cards

Campaign view

Creative ranking by status (top performer, stable, fatigued, underperformer)
Trend charts
Quick visual comparison strip

Ad detail view

Creative image + design attributes
Audience-specific effect tags (green positive, red negative)
Three audience-fit scores
Predicted fatigue/profitability-loss day with uncertainty
Recommendation: scale, watch, or pause

How we built it

Data pipeline
Loaded the core campaign and creative tables, mapped internal IDs to readable names, and cached results to keep navigation fast.

Audience-specific correlation analysis
Computed trait-to-performance relationships by device, age group, and country. We used Pearson correlation for numeric fields and signed Random Forest importance for categorical fields.

Creative fitness scoring
Combined correlation signals into three scoring styles: linear total alignment, consistency-weighted (Sharpe-style), and a focused Top-10 strongest-signals score.

Feature engineering
Built model features from first-7-day CTR/CVR/ROAS behavior and static creative attributes.

Profitability event definition
Defined profitability loss as the first day in a 3-day ROAS < 1 streak, then used that day as the regression target.

ML modeling
Trained two offline LightGBM regressors: one predicts profitability-loss day, the other predicts fatigue day. Model quality metrics are surfaced in the UI.

Dashboard implementation
Built the product in Streamlit with Plotly and session-state-driven navigation.

Challenges we ran into

Correlations needed to be audience-specific; global scores were misleading.
The dataset is perfectly uniform, so volume-based activity comparisons always tie.
Profitability-loss and fatigue are different prediction targets.
Streamlit reruns required careful state handling to preserve navigation.

What we're proud of

End-to-end flow from advertiser view to creative recommendation
Explainable, audience-specific attribute tags
Predictions shown with uncertainty bands
Fully offline execution (no API keys or external calls)

What we learned

Fatigue is multi-signal: CTR often drops before CVR.
Novelty was the strongest durability signal (about +15 days to fatigue for high-novelty creatives).