It started with skepticism, not belief
Around 2022 I got curious about astrology. Not as a believer — as someone who couldn't find anything substantial enough to believe or dismiss. So I did the field research: consulted astrologers online, bought readings from the big consultation platforms, collected reports. Some were eerily relatable. Some were completely off. Which raised the obvious question — were the hits real insight, or just general predictions that sometimes land?
I was stuck in the most uncomfortable place for an engineer: unable to fully discard it, unable to fully accept it.
The experiment that failed, and what it taught me
Then GenAI arrived. Through 2022–23 I started asking ChatGPT, Claude, and Gemini for astrological predictions. My mistake: I gave them just my birth details and expected proper chart data back. They confidently produced garbage — LLMs simply cannot compute charts.
So I changed the input. I uploaded the professionally-generated PDF reports I already had, and asked the models to interpret them. Suddenly the output was good. That was the insight the whole product is built on:
GenAI is genuinely good at interpretation, and genuinely bad at chart generation. Charts are math. Readings are language. Split the job.
The trigger: why is this even paid?
Around the same time I found a slick paid AI-astrology app doing roughly this — decent at best, charging heavily for it. Between astrologers charging per consultation and apps charging subscriptions for what is ultimately deterministic math plus AI prose, the economics offended me. The marginal cost of a reading is a few AI tokens.
So in June 2025 I started building Celestial Digits, with two explicit goals:
- Give it away free. Real Vedic numerology — Mulaank, Bhagyank, Namank, compatibility, dual-tradition daily horoscopes, downloadable PDF reports — at zero cost.
- Learn marketing on a real product. Engineering I knew. Distribution I didn't. This was the classroom.
How it's built
Next.js 15 and TypeScript, Supabase for auth and data, i18next because this audience doesn't all browse in English, PWA so it installs and survives patchy connections. Two AI providers — Gemini and the ChatGPT API — routed per task, with one as fallback when the other has a bad day.
The architectural rule is the lesson from the failed experiment, enforced strictly: the numbers always come from deterministic code; the model only ever writes about numbers it's handed. That boundary is what separates a numerology product from an AI hallucinating mysticism.
The marketing education (organic > paid, proven)
I ran the Instagram page. I ran paid ads. Ads bought awareness — but awareness didn't come back next week. Recurring users were near zero.
In 2026 I started publishing blog content on the site instead. That's what changed the curve: organic search visitors arrive with intent, read, generate a report, and return. The platform has now generated 5,000+ reports, and the recurring usage came from content, not spend. Every marketing thread I'd read said organic beats paid for retention — now I have my own receipts.
The two-day app
Daily horoscopes and AI chat sessions wanted to live on a phone, not in a browser tab. So I built the mobile app — in two days, with Claude as the pair programmer. It's live on Android. The iOS build is parked in TestFlight: App Review flags free apps in crowded categories aggressively, and untangling that is a fight I haven't had time to schedule. The goal hasn't changed — get it to people, free.
The money question
I test-ran a paid numerology report. It didn't convert — unsurprising when everything next to it is free. So today the AI tokens, the email infrastructure, and the hosting run on my own money. That's a deliberate choice, not a failure state: the bill is small, the product stays free, and the things it pays me back in — a live AI product, a marketing education, and proof that the deterministic-core/generative-surface pattern works — are worth more than the invoice.
What I took from it
- Deterministic core, generative surface. Let code do the math and the model do the prose. This is now my default pattern for every AI product.
- Organic beats paid for retention. Ads rent attention; content earns return visits.
- Free is a feature with a bill. If the bill is small enough, it's the best growth feature there is.