i'm trying to make taste measurable.
ai solved building. it didn't solve taste — and that's the only thing left that matters. i build products, then i figure out why the good ones are good.
now i'm taking that into models: what taste means for a machine, and how to benchmark it on real-world work.
United Data is neutral ground for training data: AI labs post what they need, vetted teams respond with samples, and the lab picks the best result after scoring it on their own test.
The bet is simple: the best data is not sitting in one warehouse. It is distributed across expert teams, and labs need a private, trusted way to source it without revealing strategy or managing a dozen vendor threads.
400K MAU · 10K GitHub stars
21st.dev is a reference layer for AI builders: a library of real UI patterns, components, blocks, and app screens that make generated products less generic.
Magic MCP brings those references into coding tools, so taste is available at the moment of building instead of hidden in screenshots and bookmarks.
5K GitHub stars · 600 forks in 2 weeks after launch
1Code is a visual client and workflow layer for coding agents: parallel sessions, Git review, live previews, sandboxes, and automations that keep work moving after the terminal closes.
That work expands into the Agents SDK, a production runtime for shipping CLI-based agents with auth, sandboxes, UI, and observability.
Suggesty put GPT-3 answers directly on top of Google results: the Perplexity shape before that became the obvious direction for search.
The funny receipt is that Google later featured it in an internal article about LLM-powered conversational search.
