AgenticDatingFeedback

Get prepared for the real dating game!

Inspiration

Dating apps give you no feedback loop. You post a profile and just... wait. You never learn why people swipe past — was it the first photo, a weak prompt, or that they bailed before scrolling to your best bits? People rewrite profiles blindly, guessing.

What it does

We built a usability-testing lab for dating profiles. You create a profile, and 20 AI personas — each a distinct fictional character with their own values, interests, and dealbreakers — actually browse it in real headless browsers with Stagehand, rendered on a live Hinge-style UI. Each persona sees only what's on screen as it scrolls, so a GPT-4o vision model forms a genuine first impression from real screenshots, feeds that interest score into a deterministic "attention gate," and decides to keep scrolling, drop off, pass, or like. You get a dashboard: outcome distribution, a drop-off funnel showing exactly where people lose interest, and per-persona reasoning quotes. It's progressive disclosure, just like a real swipe.

How it works

Architecture. A property-tested deterministic core (feed projection, attention gate, run controller, trace assembly) keeps the LLM out of the critical path, wrapped by an orchestrator running personas concurrently under a 180s budget, persisted to Supabase.

How we used Kiro

We built spec-first. Kiro's spec workflow turned our design into 23 tracked tasks with 32 correctness properties, and its agent + sub-agents executed them in waves — context-gatherer to map the codebase, coding sub-agents to implement and test. Kiro wrote ~205 property-based and integration tests (Hypothesis) alongside the code, which caught real bugs: a wrong Stagehand SDK method silently erroring every persona, and an .env path issue. Kiro's GitHub Power handled branches and PRs, and web fetch let us verify the live Stagehand v3 API against docs. Kiro didn't just autocomplete — it drove correctness, kept us honest with tests, and shaped a verifiable build