Building tools that think.

Hermes runs on a heartbeat, so the model bill adds up fast. I split routing by job: the interactive path I actually talk to runs on a stronger model, while the scheduled background checks drop to a smaller, low-effort one. The always-on parts got noticeably cheaper and the conversation stayed just as sharp.

The Actual Budget advisor used to nag against numbers I set once and forgot. Now it learns a baseline from my actual history and measures drift against that, so it adapts as habits change. It checks in over Telegram on a daily, weekly, and monthly cadence and calls out creeping subscriptions before they become a line item I resent.

Autopilot now sweeps the whole vault every few minutes, generating frontmatter, tags, backlinks, and freshness scores on its own, whether a note was edited by hand, in Obsidian, or by a coding agent. The whole thing is judged against one promise: the brain maintains itself. No backend, no account, and your Markdown stays yours.

The same burst of speech can now be shaped into different outputs depending on the profile you pick, so a quick thought and a structured note don't need two different tools. Everything dictated stays searchable in history and can be reused later as context, which quietly turns a dictation app into a memory.

Every morning a script reads a fresh batch of generated startup ideas, scores them on market signal, build effort, and differentiation, and surfaces the few worth a second look. For anything that clears the bar it offers to spin up a scrappy validation MVP on the spot, so the gap between "interesting" and "testable" stays hours, not weeks.

The page you're on is hand-built with almost no dependencies (three.js drives the hero; everything else is vanilla), animated with WebGL, and it ships a machine-readable view (the toggle in the nav) so an AI agent gets the same content as clean, structured text. Add structured data, an llms.txt, and an RSS feed, and it's as easy for a crawler to read correctly as it is for a person.