CASE STUDY
Specialty CPG portfolio
AI-powered incrementality intelligence across a multi-brand consumer portfolio
THE SITUATION
Marketing decisions running on rolled-up vendor reports
A specialty CPG group runs multiple consumer brands across a regulated, retail-heavy market. Trade promotions, in-store discounts, sampling programs, and seasonal campaigns all compete for the same dollar. Decisions on where that dollar should go were being made on rolled-up monthly numbers and vendor-supplied scorecards: directionally fine, causally meaningless.
Lift was being claimed by every channel and every vendor. The team needed a defensible answer to a single question: which dollar actually moved the needle, and which one was paying for sales that were going to happen anyway.
BEFORE
Marketing decisions made on rolled-up sales numbers and vendor scorecards.
AFTER
Causal lift behind every promo dollar. Real experiments designed in days, not weeks.
WHAT I BUILT
An incrementality intelligence engine over the entire portfolio
I built a causal incrementality platform spanning every brand, every campaign, every retail channel. Time-series models reconstruct the counterfactual baseline for each promo window, so lift only counts if it is statistically distinguishable from what would have happened anyway.
On top, an AI-assisted experiment design layer: sample-size and power analysis baked in, automated A/B and geo-test recommendations, and an evaluation surface that scores every model release against held-out historical campaigns before it goes live. Anomaly detection on retail performance flags unexpected behaviour with attribution back to the campaigns running underneath.
Causal lift, not vanity lift
Counterfactual baselines built per brand, channel, and window. Real lift only counts if it is statistically distinguishable from what would have happened anyway.
AI-assisted experiment design
Sample size, power analysis, and geo-test recommendations automated. Marketing leads design rigorous tests without an analyst in the loop.
Forecast and anomaly detection
Time-series models on retail performance, with anomaly flags when a brand or store diverges from forecast. Attribution back to the underlying campaigns runs alongside.
THE OUTCOMES
Marketing dollars rebalanced on causal evidence
Causal
lift attribution behind every promo dollar
Days
to design a rigorous experiment, instead of weeks
Live
rollout across every brand in the portfolio
STACK
- Python
- Postgres
- OpenAI
- Time-series
- Causal inference
- Evals
