What happens when you combine modern AI with ancient fortune telling traditions? You get Tanya Peramal (tanyaperamal.com) -- a platform where users consult an AI Oracle powered by Groq's lightning-fast inference for Ba Zi, Tarot, I Ching, and Gui Bu readings.
The brief sounded simple on paper: build a fortune telling website. In practice it sat at the intersection of four different disciplines -- traditional divination knowledge, conversational AI prompt design, real-time backend engineering, and SEO for one of the most competitive niches in Indonesia. This article breaks down the architecture, the design decisions that actually mattered, and the mistakes we would not repeat on the next AI product.
The Problem We Were Solving
Most 'AI fortune teller' demos online are a thin wrapper around a chat model: a text box, a generic system prompt, and a wall of text back. That works fine for a novelty page, but it falls apart as a product. The Tanya Peramal brief called for something that felt like consulting an actual practitioner -- fast, in character, with real domain knowledge of Ba Zi, Tarot, I Ching, and Gui Bu -- and commercially sustainable through a token system instead of banner ads.
The Tech Stack
Frontend: React + TypeScript + Tailwind CSS
The frontend is a standard React SPA built with Vite for fast builds and hot reload during development. We styled it with Tailwind CSS using a mystical dark purple and gold palette that fits the fortune telling theme without tipping into cliche. The site has grown to 62 pages covering the different divination methods, zodiac content, and educational articles, all sharing the same component library so a new page can ship in an afternoon.
Backend: Express.js + SQLite
The backend runs on Express.js with SQLite for user data and token management. We chose SQLite over PostgreSQL deliberately: the data model is simple (users, tokens, sessions), traffic is read-heavy rather than write-heavy, and a single-file database means one less server to patch, monitor, and pay for. If the platform outgrows SQLite's write throughput later, migrating to PostgreSQL is a known, well-documented path -- we did not want to pay that complexity tax on day one.
AI Engine: Why Groq Over a Direct OpenAI or Anthropic Call
For the AI Oracle itself, we route through Groq as the primary provider, with OpenRouter as a fallback. Groq runs open models on custom LPU hardware instead of GPUs, and the difference shows up directly in the product: where a typical GPT or Claude call can take 4-8 seconds for a long response, Groq streams the same length in under 2 seconds. For most SaaS features that gap does not matter much. For an oracle, it is the entire experience -- a fortune teller who makes you stare at a loading spinner for ten seconds stops feeling mystical and starts feeling like a broken app. OpenRouter sits behind Groq as a fallback: if Groq rate-limits during a traffic spike or has an outage, the request routes to an equivalent model through OpenRouter instead of the user seeing an error.
System Prompt Engineering
The most critical part of the project was the system prompt, not the frontend. The model needs to respond as a mystical fortune teller, not a helpful assistant with bullet points and disclaimers. Here is what we focused on:
We rewrote the prompt more than a dozen times, mostly by reading real transcripts and flagging any line that sounded like a chatbot rather than a fortune teller. A phrase like 'Based on the information provided' had to go; 'The stars whisper of a change coming in the second month' had to stay. That kind of editing does not show up in a diff of lines changed, but it is the difference between a product people trust and one they close after one message.
Token Credit System Design
We implemented a credit-based system instead of a subscription. Subscriptions make sense for daily-use tools; fortune telling is closer to an occasional, high-intent purchase, so pay-per-use fits the actual usage pattern better. Here is how it works:
This is simpler to build and audit than a full subscription/billing integration, and it gives users direct control over their spending instead of a recurring charge they forget about. The Google OAuth integration also means zero password reset flows, zero forgotten-password support tickets, and one less place for credentials to leak.
Session Management: Progressive Answer Depth
Each oracle session allows up to 3 follow-up questions within a 5-minute window, and the answer depth decreases progressively:
This design does two things at once. It nudges users to ask their most important question first instead of drip-feeding trivial follow-ups, and it naturally caps AI compute cost per session without a hard word limit that would feel arbitrary to the user. After 5 minutes or 3 questions, the session expires and a new token is required to start again.
Keyword Filtering for Safety
Before any query reaches the model, it passes through a keyword filter that checks for medical terms (symptoms, diseases, medication names), financial terms (stock tickers, specific investment advice), legal terms (lawsuits, contract disputes), and harmful content (self-harm, violence). Blocked queries return a gentle in-character redirect -- the Oracle explains it cannot see that domain of the future and suggests consulting a professional instead. That framing matters: a hard 'This request violates our policy' message breaks immersion and reads like a corporate bot, which is exactly what the persona is built to avoid.
Divination Methods We Had to Get Right
None of these methods can be faked with generic prompting. Each one has real rules, and users who know the tradition will notice immediately if the AI gets them wrong.
Ba Zi (Four Pillars of Destiny)
Ba Zi maps a birth date and time to four pairs of Heavenly Stems and Earthly Branches, then reads the balance of the five elements across them. Getting the pillar calculation right is a deterministic date/time problem, not something to leave to the language model's arithmetic -- we compute the pillars in code and pass the structured result into the prompt, so the AI only has to interpret, not calculate.
Tarot
The Oracle draws from a full 78-card deck, tracks upright versus reversed orientation, and respects spread position (past/present/future, or a Celtic Cross layout) when interpreting a card. The interpretation changes meaningfully depending on which position a card lands in, and the prompt has to carry that context explicitly.
I Ching
We simulate the traditional three-coin toss method to generate one of the 64 hexagrams, including any changing lines that produce a second, transforming hexagram. The AI then interprets the hexagram text and the changing lines together, which is closer to how a real I Ching reading works than just picking a hexagram at random and summarizing it.
Gui Bu (Turtle Shell Divination)
This is the method with the thinnest coverage in general AI training data, since it is a much older and less globally documented practice than Tarot or Ba Zi. We had to source additional reference material on the crack patterns and their traditional interpretations and fold that into the domain knowledge section of the prompt so the Oracle does not just improvise when a user picks this method.
SEO for a Fortune Telling Site
Fortune telling is a competitive niche in Indonesia. We optimized for long-tail keywords like 'ramalan online gratis', 'kalkulator Ba Zi Indonesia', and 'tarot reading AI' rather than fighting for the single highest-volume head term. Each of the 62 pages has a unique meta title, description, and schema markup, and we implemented FAQ schema for common questions about each divination method, which helped several pages earn featured snippets within the first few weeks of indexing.
Keeping the Oracle Fast Under Real Traffic
Sub-2-second responses on a quiet server are easy. Keeping that number steady when multiple users hit the Oracle at once required a few deliberate choices: streaming the AI response token by token instead of waiting for the full completion, so the user sees the reading appear progressively rather than staring at a blank screen; setting an aggressive timeout on the Groq call so a slow request fails over to OpenRouter instead of hanging; and keeping the system prompt itself lean, since every token in the prompt is a token of latency before the model starts generating the parts the user actually reads.
Lessons Learned
1. AI persona consistency matters more than feature count
Users notice within one message if the Oracle breaks character. We spent more time rewriting the system prompt than building the entire frontend, and that time was well spent.
2. Speed is non-negotiable for this category of product
A fortune teller that makes you wait 10 seconds loses all mystique before it says a word. Groq's sub-2-second responses are not a nice-to-have here, they are the product.
3. Simple auth wins
Google OAuth eliminated the majority of auth-related support requests compared to email/password registration, and it removed an entire category of security surface (password storage, reset flows, credential stuffing) we did not have to build defenses against.
4. Deterministic calculation plus AI interpretation beats AI doing everything
Anywhere the method has hard rules -- Ba Zi pillars, I Ching hexagram generation, card draws -- we compute the deterministic part in code and hand the AI a clean, structured result to interpret. Letting the model both calculate and interpret invited small arithmetic errors that undermined trust with anyone who actually knew the method.
Conclusion
Building an AI fortune teller taught us that the best AI products are not about the technology -- they are about the experience. The React frontend, Express backend, and Groq inference are just tools. What makes Tanya Peramal work is the carefully crafted persona, the progressive session design, the deterministic-plus-AI split for each divination method, and the respect for traditional divination practices most AI wrappers ignore.
Building something similar for your own niche? [Get a free quote](/get-quote/) and we will map out what an AI-powered product would look like for your business.