Overview
An AI virtual astrologer that generates personalised daily horoscopes from birth-chart data, delivers Tarot and Oracle readings via LLM + RAG, and guides users through the platform with a gamified conversational assistant.
The problem
The platform needed daily, personalised astrological content and readings at scale, plus an engaging way to guide users across its modules — without generic, repetitive LLM output.
Context
A consumer-facing US product where personalisation, tone consistency and domain accuracy directly drive engagement and retention.
Constraints & challenges
- 01Predictions must be personalised to birth-chart inputs (date, time, coordinates), not generic.
- 02Domain-specific Tarot and Oracle knowledge has to ground the LLM output.
- 03Conversations should remember user context across a session.
- 04Guidance across modules must feel like a game, not a help menu.
My role
Built the LLM and RAG layer: personalised prediction generation, retrieval over astrology knowledge for Tarot and Oracle readings, the gamified conversational assistant, and vector-based contextual memory.
The solution
Combined structured birth-chart data with prompt engineering to generate daily predictions, used a RAG layer over curated astrology and Tarot/Oracle content in Pinecone to ground readings, and implemented vector embeddings plus session memory so the conversational assistant stays personalised and context-aware while steering users through modules with reward-based engagement.
- Personalised daily horoscopes from birth-chart data
- Tarot & Oracle readings grounded with RAG
- Gamified conversational assistant
- Session-aware contextual memory
System architecture
System architecture
- Source
- Storage
- Processing
- Model / LLM
- Agent
- Interface
Two grounding sources feed the LLM: structured birth-chart data through prompt templates, and retrieved domain content from the vector store. Session memory closes the loop for continuity.
Key engineering decisions
- 01
RAG over curated domain content
Grounding Tarot and Oracle readings in curated knowledge keeps output accurate to the practice and less repetitive.
- 02
Vector-based session memory
Embedding earlier turns lets the assistant stay personalised across a session without stuffing the whole history into every prompt.
- 03
Gamified guidance layer
Reward-based journeys keep users exploring modules — engagement design is part of the AI system.
Technology
- Python
- LangChain
- Pinecone
- LLM APIs
- RAG
- Prompt Engineering
- Vector Embeddings
Results
Delivered personalised, session-aware astrological content and readings with a conversational assistant that guides users across the platform through reward-based engagement.
Lessons learned
- Consumer LLM products live or die on tone consistency and personalisation.
- Memory design is a retrieval problem.
Building something similar?