From 2021 to 2024 I worked on measurement and identity for cross-screen advertising — the statistics of knowing who saw what, and what it did, when your data is enormous, fragmented, and incomplete.
Each of these took a research-grade statistical method and made it run reliably over hundreds of millions of records in production.
120M households · 10M points · ~1M params
Designed and implemented a Gibbs sampler for Bayesian non-parametric density estimation to fill missing demographics in a deeply nested household structure — tuned to run over 10 million data points and roughly a million parameters.
100M nodes · 1B edges
Built a probabilistic framework for the evolution of interconnected bipartite graphs, linking devices, identities and demographics from multiple sources over time, with an emphasis on geographical consistency.
100K+ time series · zero-inflated · autoregressive
Modeled purchase and TV-viewership as household-level Poisson point processes, separating the true event-generating rate from selection and masking effects with a zero-inflated, hierarchical model that shares information across households. Covariate link functions and autoregression let each household's intensity depend on its attributes, on time, and on its own recent history — across 100,000+ time series.
N(X,Y) = Nₛ(X,Y) / p(S=1 | X,Y)
Built a projection framework for unbiased measurement of household–event pairs under incomplete observation. From an exchangeability argument it recovers any population quantity by dividing the observed count by its probability of being observed — inverse-probability weights set by identity-graph matchability and observability — while estimating the total universe intensity jointly. It turns a biased, partially observed panel into calibrated, unbiased ad-effectiveness estimates.
