Dash

01 / THE SYSTEM

One state layer for everything

Data inputs to the system flow through server-owned adapters where they are normalized into structured records in Postgres, then they can be surfaced through Home, Tasks, Health, Finance, Review, or queried through Brain. I kept as many calculations, rankings and canonical records deterministic, with AI models being called for fuzzy inputs such as classification, embeddings, or search.

Importantly, the models can interpret data, but not edit the original inputs.

I started out by hosting the system and scheduled jobs in the cloud and eventually started moving the heavier workflows toward a private always-on worker.

02 / STRUCTURED STATE

Turning messy inputs into something I could trust

I imagined the system as being able to speak into my phone anytime I wanted and say something like "I need to finish the mechanics problem set Friday" or "Just met Costanza from IBM" and it would automatically go where it belonged.

The model pulled structure out of the sentence—what it was, when it was due, what it related to—then Dash would store the hard deadline, decide the day I should focus on it and its current priority. That meant ordinary code would rank the queue, and if a capture was uncertain, it would route to Review instead of silently committing it.

03 / MEMORY + AUTONOMY

Useful AI needs boundaries

The Brain is my attempt at building a vector DB of my life. I embedded sanitized summaries of existing state into pgvector, then kept references back to the original records, which allowed me to create a chat where answers can cite the original input they were using.

04 / PRODUCTION

The one-minute sync that wrecked the database

I had about twenty calendars to keep fresh and Google's API quotas made a one-minute sync look cheap.

Turns out I was measuring the wrong system.

There was a huge bottleneck downstream, where scheduled functions started overlapping, making Calendar consume roughly 83% of the system's entire scheduled runtime while the database was doing an absurd amount of repeated work: roughly 47.6 million calendar_events updates against only ~4.3K live rows.

So I shut the sync down, let Supabase recover, dropped the cadence to every fifteen minutes, and now I know not to size a system against just the API quota.

05 / WHAT I LEARNED

I added integrations faster than I defined to the existing system what they meant

Looking back, this was clearly the recurring mistake. I kept adding APIs because it was easy, but then failed to immediately define to the system whether their data was authoritative, whether a repeated value was an event or a snapshot, how to reconcile conflicts, and what to do if that part of the pipeline failed.