Statistics at the point of measurement.

Currently in development for a Utrecht University sociologist where the underlying question was, "How do we measure exposure to crime-related videos on TikTok?" Here too we're working in constrained compute environments in service of privacy: we need modern methods to answer these questions at scale, but we have a responsibility to be careful with respondents' donated data.

For this, I developed a Rust pipeline that takes donated TikTok exports, fetches the videos the respondents watched, transcribes and classifies them with local models, at the scale of 1 million videos. Viewing history never leaves the secure environment to be processed by a third-party. The engineering serves to get everything working, but the remaining questions are statistical: how do you classify transcription or classifier uncertainty into the resulting estimates so that we can provide an honest inference?

A comparative study of API-based data access versus DSAR-export access for data-donation research across major platforms. Moving from one-off exports to APIs reshapes the measurement system itself: it relocates computation, opens up longitudinal rather than one-shot observation, and changes the privacy guarantee of the pipeline. The part I care about is the statistical one: what that does to validity, representation, and the kinds of inference these systems can actually support.