← Back to Vivid Labs

Case Studies

MeasurementDisney StreamingPerformance marketing analytics

Disney+ goes to infinity and beyond

Disney+ was marketing across Snapchat, Facebook, TikTok, paid search, and brand for every title release and product launch. Each channel team had its own reporting and its own definition of success, so the honest answer to "where should the next million go?" was a negotiation rather than an analysis.

The fix wasn't a better dashboard. It was a model that put every channel on the same scale.

  1. Adjusted spend inputs for diminishing returns so the regression stopped rewarding channels for being saturated.
  2. Fit a multiple linear regression media mix model across channels, with macroeconomic controls, and supported the rollout with the measurement team.
  3. Translated the model output into a concrete spend recommendation for the Disney+ Day campaign.

What made it stick: executives could see spend recommendations in a time where cookie deprecation was impacting meausrement reliability.

Applied AIGoogleData center compliance

Google automates repetative tasks

Compliance and business requirements lived in PDFs, decks, and spreadsheets in a dozen inconsistent formats. Extracting them by hand was slow, inconsistent between reviewers, and difficult to audit.

We built an agentic system to automate the extraction of these requirements. The interesting engineering wasn't the extraction but everything built around it that made the output trustworthy.

  1. Built the core agentic extraction pipeline on Gemini 3.1 Pro through the google/genai SDK, tuned to over-extract rather than miss a requirement.
  2. Added AI-driven verbatim verification: every extracted requirement is checked back against the source text, so anything the model invented fails the check instead of reaching a reviewer.
  3. Redacted PII locally with regex on the server before any content reached the model API, and kept API keys strictly in server-side environment variables.
  4. Made it survive real traffic with exponential backoff and jitter via async-retry, absorbing rate limits and transient failures instead of dropping documents.
  5. Mapped validated requirements to predefined industry domains and common control IDs, persisted to BigQuery so the results were queryable alongside everything else.
  6. Shipped the whole thing as a React 18 and TypeScript single-page app on a Node and Express backend with a lightweight SQLite store.

What made it stick: an unbroken trail from source, to extracted requirement. A reviewer can always ask "where did this come from?" and get an answer.

Have a measurement or AI problem like these?

Thirty minutes, no charge. Tell us what you're trying to measure or automate and we'll see if we can help.

Book a free 30-minute call