Screenshots are awesome. They capture everything we actually do online: what we want, where we've been, what made us laugh. At Seenit, I designed and built a system that turned those screenshots into AI-generated statuses for friends. In a 10-day beta, 22 users generated 705 POVs.

In beta, only 30% of cards were being shared. I identified recurring patterns of these cards, and built a testing sandbox to isolate and manipulate 8 core model attributes, which allowed for rapid iteration of a new system prompt layer.

I designed a rubric scoring gate, but found that a model scoring its own outputs is inherently biased. I introduced an independent judge model to evaluate candidates comparatively rather than in isolation. The model's second attempt consistently outperformed the first.

When the AI gets it wrong, it breaks trust. A single tap lets users ask the model for a different interpretation rather than editing or dismissing. This kept misses low-stakes, and turns a bad AI hallucination into a fun interaction rather than a product failure.

I opted for a pipeline that prioritised output quality and cost over speed, which meant building around the constraint. In addition to a processing experience in-app, I built a push notification and live activity flow so users could share without opening the app or composing anything.

1. In a 10-day beta, 22 users generated 705 POVs. This meant the habit was forming before distribution was solved
2. A model scoring its own outputs is inherently biased, it rates what it generated highly because it generated it.
3. Presence > identity. Friends care more about what you're in the middle of than how you present yourself