The AI underneath — Astral Studio

What it does

Quietly. In your voice.

Drafts your launches

When you announce a new retreat, the first draft of the email lands in your inbox. Written in your phrasing, structured the way you usually structure these announcements. Edit it down. Send it. Twenty minutes saved per launch.

Answers a student at 3am

In the language of your tradition. The terminology you use, the lineage you teach, the depth your students expect. It escalates anything subtle to you — but it handles the questions you've answered a hundred times.

Sorts incoming intake forms

Reads each one, tags it, flags the ones you'll want to handle yourself. The 9 forms a week become 3 you actually read, with notes from the AI on what to look for in each.

Learns every time you teach

Your live recordings get transcribed and folded into the model. Three months in, the AI has heard everything you've said in every cohort. It speaks the way you speak — because it learned from you.

What it's trained on

Your work. Not the internet's.

Generic AI knows what the internet has decided about your tradition. Your AI knows what you have decided about your tradition. Different thing.

Past emailsYears of how you actually answer students, clients, and inquiries

Intake formsThe questions you ask + the answers you respond well to

Live recordings + transcriptsEvery session, every cohort call, every retreat dharma talk — folded in as it happens

Written workBooks, blog posts, course curriculum, hand-written notes you scan in

Decision historyThe students you accepted vs. declined, the offerings you launched vs. shelved — patterns the AI picks up over time

Where it lives

On your infrastructure.

Not a SaaS subscription to my AI. Your model, your weights, your prompts — running on a server in your name. If you ever stop working with me, the AI doesn't leave. It can't, structurally.

For most clients this means a fine-tuned base model (OpenAI, Anthropic, or local depending on data sensitivity) with a vector store of your training material. The data never leaves your account.

What you own

→The fine-tuned model weights

→The vector store of your content

→The prompts + system instructions

→The conversation history

→The integration code

→The API keys (in your name)

Standard tools. Migrate-able. No proprietary lock-in.

What it's not

Not that.

×A chatbot widget in the corner of your homepage

Those are decorative. The AI here doesn't need a UI — it works underneath, in your email, in your forms, in your inbox.

×A generic ChatGPT wrapper

A wrapper trained on the internet talks like the internet. Yours talks like you because it learned from you.

×A SaaS subscription you can't cancel

No vendor between you and the model. Your weights, your account, your bill.

×A magic feature in a marketing deck

It's software. It does specific things well — drafts, answers, sorts, learns. It doesn't replace you. It carries the boring weight you carry now.

The honest part

It's not magic. It compounds.

Models plus your data plus good design. That's the whole formula. There's no proprietary breakthrough — the breakthrough is that someone bothered to build it around the specific shape of your work.

Month one, the AI is decent. It writes drafts that sound mostly like you. Students get useful but slightly off-key answers. Forms get sorted, sometimes wrongly.

Month six, it's uncanny. The drafts barely need editing. Students reply with "thank you" not knowing it wasn't you. Forms come pre-sorted with notes you'd have written yourself.

Month eighteen, it's irreplaceable. Years of your teaching folded into a model that nobody else can replicate — because nobody else has your training data.

That's why I build for the long term. The AI gets useful slowly, and then suddenly very useful. The compounding starts on day one.

Curious if your work has enough material to train on

Send me a sample of your writing.

Three to five emails you've written to clients, an intake form you use, a session note you took. Twenty minutes of reading and I can tell you whether the AI substrate fits your work or not. No pitch, no pressure.