Show HN: Potential conversation partners in Bluesky using agentic search (sketch.wsng.app)

What if instead of talking to AI, we could find human counterparts currently sharing the same interest to talk to? Some whales kind of do that by singing across the ocean.

This inspired a thought process for a conceptual social network - whalesong - where an idea, thought or a conversationally valuable statement could be matched on.

In contrast with the wide-broadcast social network visibility (restricted on groups or connections), I’d love to explore the idea of context-based networking.

This experiment is showcasing the technical feasibility and potential value in whalesong.

The goal: uncover conversational counterparts among Bluesky posts using an agentic search algorithm.

The setup: a one-week window of bluesky posts with some filtering (e.g. mostly text, no promotions or reposts/comments), indexed in a postgres database. A curated seed set - an agentic process defines a conversation-seeking cohort by creating some heuristic queries, iterating and validating it on retrieved data and resulting in a non-exhaustive set of high-quality data points. Counterpart retrieval - for a seed conversation-seeking post, an agentic process generates search queries, runs them against the database and retrieves potential matches. It reviews retrievals and iterates on improving or expanding the search.

Out of several configurations, the best performing agentic setup (bespoke mini-harness with Gemini Argon - low) managed to find meaningful matches to 79% of a sample set of 100 seeds. This setup consumed on average 42K input tokens on each match search, of which 56.8% were cached, output token count was negligible.

So this experiment made it clearer to me that technology is not the blocker for a “whalesong” experience. This leads me to further investigate the experience itself. Would love to hear from you folks about experiences in getting in touch with a complete stranger on the web because both of you asked the same question or had a similar problem to solve. What made you connect? Where did it happen? Did finding similarity make connecting easy?