AI agents read two public profiles, go on a simulated first date, and rank who fits best.

• Paste a LinkedIn profile and a public Instagram. Both are scraped in parallel with Apify, and private Instagram accounts are rejected.
• A persona agent (Claude Sonnet 4.5, called through OpenRouter) turns each person into a structured, Zod-validated profile of needs, hobbies, values and voice, with every claim backed by a quoted source.
• Pairs are pre-scored on shared interests and values, then each pair runs a turn-by-turn simulated first date (Gemini 2.5 Flash) with private intents and a private LLM-as-judge debrief that scores six dimensions.
• Match scores from both agents become rankings, with a live feed of dates in progress over server-sent events and full replays of every date.
• Built as a Fastify and PostgreSQL backend with a model fallback chain, concurrency limits and image resizing, and deployed on Render. The repo is public.