astralis

Closed beta · Independent research project

A quieter place to think about work.

Astralis is a reflection companion for burnout and work stress. It grounds what it says in published research rather than its own confidence, keeps track of your thread between conversations, and connects you with other people using it — under whatever name you choose.

Begin with your emailNo account elsewhere needed. Nobody hereever sees your email — only a name you choose.

How it works

retrieval

It reads the literature

A scheduled job pulls peer-reviewed work on occupational stress from OpenAlex, chunks it, embeds it into pgvector. The library grows weekly. No model is retrained.

memory

It remembers your thread

A short factual summary per session, retrieved as context next time. Rows in Postgres — auditable, and deletable in one click.

feedback

It learns what helped

Mark a reply useful and the passages behind it are reweighted for retrieval. That reweighting is the whole of its self-improvement.

safety

It knows its limits

A rules-plus-classifier filter runs before the model. Flagged messages never reach it, never enter retrieval, never enter memory — they get a fixed card written by a person.

What it is not

  • Not therapy, not diagnosis, not a clinical tool of any kind.
  • Not a crisis service. Crisis language is intercepted before the model sees it.
  • Not anonymous to the operator: messages are stored, and processed by Anthropic's API.
  • Not a product. There is no company, no pricing page, and no plan to build one.