// journal

Notes & signals

  • Brain-first: the memory every agent checks before it answers

    Agent sessions start from zero and re-derive the same decisions endlessly. GBrain is the self-hosted memory server every agent I run checks before it answers, and writes back to after it works.

  • Three eras, one method

    Why RandomSynergy isn’t an AI studio, and what two prior technology cycles taught me about this one.

  • One week, 1,189 tests: building The Foundry cockpit

    My portfolio intelligence kept freezing in static handover docs, so I built a system that keeps it live: profiled, scored, and routed to a decision. It took one week and 1,189 tests.

  • Deterministic first, LLM second

    The reliable way to ship an LLM feature is to make the model the last layer, not the first. Here’s how MeetKlay’s diagnosis is built.

  • Playwright can’t tap: giving an agent a real iOS device

    CI runs Playwright, which renders Chromium, not the WebKit an iPhone actually uses. iPhone-only bugs never fail where you’d catch them, so I built an engine that drives the real device.

  • Generation belongs in an MCP server

    An agent shouldn’t break flow to open a design tool when it needs an image. Generation should be a service any agent can call, which is what I built RNSNB to be.