Client Impact & Projects

Leadership was scaling Google Display spend based on a reported 2.9x ROAS. I designed a matched-market geo-holdout experiment across 16 DMAs using a difference-in-differences framework. The true incremental ROAS was 1.2x — a 58% overstatement. Paired with Bayesian MMM built in PyMC, I projected an 18.4% revenue uplift from budget reallocation without adding a single dollar of total spend. This changed how the company made every subsequent media investment decision.

Last-click attribution was starving mid-funnel channels of credit and distorting Q1 budget decisions. I rebuilt the attribution model using a Markov chain framework, calculating removal effects across the full customer journey on Google, Meta, and Amazon. The result: $34.2K in misallocated budget surfaced, mid-funnel channels properly valued, and a Q2 spend strategy reoriented around true channel contribution rather than platform-reported last-click.

Weekly reporting was consuming 4.5 hours of analyst time on repetitive synthesis, formatting, and packaging before insights could reach decision-makers. I built an agentic AI copilot using the Claude API, Python, and Streamlit that ingests raw performance data, applies structured analytical reasoning, and outputs decision-ready executive narratives — compressing the entire workflow to 45 seconds. This is analyst leverage that freed capacity for higher-order measurement work.

The business was losing subscribers without a reliable early-warning system. I built a churn prediction model that identified missed payments as the single strongest signal — 5x more predictive than any demographic attribute. The model output was translated into practical retention triggers and prioritized action lists, driving an 18% reduction in churn with more targeted, higher-confidence intervention.

Subscriber drop-off was being misdiagnosed as a targeting problem. RFM analysis revealed the real issue — a $342 vs $38 LTV gap between Champions and at-risk segments, and a nurture flow that was not capitalizing on high-value behavior signals. I redesigned Klaviyo flows, shifted paid budget toward Champions lookalikes, and realigned send-timing through cohort repurchase window analysis — producing 4.8x Instagram ROAS and a 30% online sales lift.