We Built an AI Astrologer That Actually Remembers You. Here's What It Took.

What we shipped

Origo is an AI astrology companion for Android. Her name is Vega. She holds your full natal chart in context, reads tarot through the transits happening in your sky right now, and remembers every reading across sessions.

Origo is live on Google Play →

That last part, the memory, is the whole reason we built it. Every other app in the category either pushes content at you (Co-Star, daily horoscopes, no follow-up possible) or fakes personalization on top of a generic engine. Vega carries the thread from last week into today. That is not a feature. It is the product.

Why this category, why now

There are roughly 30% of US adults who consult astrology, tarot, or similar at least once a year. Among women 18-49, that number is 43%. This is not a fringe thing. It is a stable, decade-long behavior that has quietly built a market worth several billion dollars a year.

The dominant apps in that market are not well-liked by the people who actually pay for them. Co-Star is beautiful and one-directional: it speaks, you listen, there is no thread. The Pattern added a conversational feature recently, which is directionally right. Nebula has a well-documented billing problem: the appnebula.co entity sits at 1.5/5 on Trustpilot, driven almost entirely by complaints about a low intro price that quietly becomes a much larger charge. (We verified this across complaint trackers and App Store community threads before writing a word of copy about it.)

The gap we saw: nobody had built a companion that (1) uses your actual chart math, not sun-sign generic copy, (2) maintains a real conversational memory, and (3) bills you like a product that respects you.

The memory wedge: what it actually means technically

Vega is backed by Gemini 2.5 Flash running inside Supabase Edge Functions. The birth chart is computed server-side using circular-natal-horoscope-js, real Swiss Ephemeris math, validated against the live sky. That JSON gets injected into every prompt.

The memory works because every time a user answers one of Vega's questions, that answer gets persisted as a memory_fact row in Postgres. On the next conversation, we pull the relevant facts back into the prompt. It is not particularly exotic as a system. The point is that it is there, and competitors structurally cannot reach it: Co-Star is a content delivery pipeline, CHANI is human-written and will not use AI, Nebula's "AI" is a thin wrapper that forgets you immediately.

The moat is not the model. The moat is the accumulated reading history that only exists inside Origo.

The billing-transparency moat

We decided early that the billing story was as important as the product story. Not because we are morally superior to other apps. Because it converts.

A large, burned, design-literate segment of our target audience has had a specific bad experience: they signed up for a "reading" at a low price and got charged a much larger subscription fee they did not clearly agree to. They are actively searching for an alternative. The message they respond to is not "better features," it is "we will not do that to you."

So Origo shows the cancel button on the paywall. Sends a reminder two days before renewal. Honors cancellations immediately. No countdown timers, no "upgrade to see the warning" manufactured fear. The FTC has been active on exactly these practices in 2025-2026.

That design choice is both the ethical thing and the higher-converting thing on this particular audience. We did not have to choose between them.

The stack

• Expo SDK 56 + Expo Router (file-based nav)
• Supabase (Postgres + Edge Functions + auth + Row Level Security)
• RevenueCat (entitlements, paywall rendering, webhooks)
• Superwall (paywall A/B layer, remote config)
• Gemini 2.5 Flash (Vega's brain, server-side, key never in the client)
• PostHog (product analytics + session replay)
• circular-natal-horoscope-js (chart math, runs in Deno Edge)
• EAS cloud builds (never local gradle)

The AI edge function pattern came out of prior work on a journaling app. The chart computation was the part that needed the most vetting: we tested three libraries and two of them could not run in Deno Edge. The one that does is circular-natal-horoscope-js under the Unlicense.

What we got wrong

The analytics gap. We shipped v1 with zero product analytics wired in. PostHog was added on a separate pass before the Play submission. This is a known bad pattern and we still did it. If you are building a similar app: instrument your core funnels before you submit, not after.

The billing propagation test. RevenueCat's Play billing connection requires a purchase test on a fresh track. We caught a "not configured for billing" error late and had to fix and re-test. The entitlement audit saved us here: we ran it systematically before submission.

Where it is now

Live on Google Play as of July 2026: download Origo. If you are building in the consumer AI space and want to compare notes on the chart-in-context approach or the billing design, reply here or find me on X.