Behind the scenes of how we build tools that turn ideas into scroll-stopping creative. Engineering deep-dives, system design, and the occasional war story.
When processing thousands of images from multiple sources, how do you prevent duplicates without relying on URLs or filenames? The same image might come from Airtable, user uploads, or external APIs - each with different URLs and metadata.
Firebase Data Connect's @auth(level: USER) directive requires authenticated user context, but Cloud Functions run in a server context with Admin SDK privileges. Simply verifying a token doesn't provide that context to Data Connect.
When syncing 100k+ records from Airtable, what happens if your function times out halfway through? Traditional ETL jobs either succeed completely or fail completely - losing all progress on failure.
When AI models fail (and they will), your users shouldn't notice. Rate limits hit, networks timeout, providers deprecate models, and services go down. A single-model dependency creates a Single Point of Failure (SPOF) that can bring down your entire feature.
How do you build semantic search that handles 100k+ assets in under 50ms? Traditional keyword search falls short when users search for concepts like 'vibrant summer colors' or 'minimalist tech aesthetic' - queries that require understanding intent, not just matching text.