Work in progress
Experiments are prototypes, proof-of-concepts, and half-built ideas. Some will ship as products. Others teach us what not to do. All of them move us forward.
Can a lightweight classifier determine which AI agent should handle a given task? Testing accuracy, latency, and fallback behaviour in a multi-model pipeline. Early results show 87% routing accuracy with sub-50ms overhead.
A renderer that displays AI-generated markdown as it streams, without layout shift. The challenge is parsing partial markdown correctly while keeping the DOM stable. Currently handling 94% of edge cases in our test suite.
A build tool that takes a single JSON token definition and outputs CSS custom properties, Tailwind config, and Figma variables. The goal is one source of truth for all brand tokens across platforms.
Moving preview authentication from API routes to edge functions for lower latency. Testing cookie handling, secret rotation, and session management across Vercel edge runtime. Focus on maintaining security while cutting response time by 60%.
How we experiment
Question
Start with a specific, answerable question. No vague exploration.
Prototype
Build the smallest thing that could answer the question. Days, not weeks.
Validate
Run it against real data. Measure. Compare to the baseline.
Decide
Ship it, iterate, or kill it. No zombie experiments.