Most AI astrology products begin with a prompt.
We chose a different architecture for OpenFate: calculate the chart first, freeze the structured evidence, and only then ask an AI model to explain it.
The problem with prompt-first generation
Explain the technical problems you observed:
- Language models should not calculate calendrical or astronomical data
- The same input can produce inconsistent chart facts
- Persuasive prose can hide incorrect calculations
- Regenerating a report can silently change its underlying claims
Include one concrete example you personally encountered.
The calculation-first pipeline
Explain OpenFate’s architecture:
- Validate birth data, location, time zone, and user options
- Run the appropriate deterministic chart engine
- Convert the result into structured facts and evidence
- Give only that evidence to the AI interpretation layer
- Validate the returned report
- Save the completed report as a frozen snapshot
Add a simplified TypeScript example based on the real implementation.
Why separate calculation from interpretation?
Cover the practical engineering benefits:
- Reproducible chart facts
- Easier debugging
- Clearer AI boundaries
- Better automated testing
- Safer retries
- Existing reports do not silently change
- Different AI models can explain the same evidence
Supporting multiple systems
Describe the engineering challenge of supporting Bazi, Zi Wei Dou Shu, Western astrology, Vedic astrology, Human Design, Sukuyo, and the six oracle systems without pretending they all share one calculation model.
Explain what is shared and what remains system-specific.
What did not work
Share two or three real mistakes, failed approaches, or architectural decisions you changed.
This section is important: DEV readers respond better to genuine engineering lessons than polished product promotion.
What we learned
Summarize the reusable lessons for developers building evidence-grounded AI products.
The approach applies beyond astrology—to finance, legal research, healthcare information, education, and any product where generated prose must remain grounded in structured source data.
Try the architecture
OpenFate is the product where we are applying this calculation-first approach:
I would be interested to hear how other developers separate deterministic computation from AI interpretation in their applications.
Disclosure: This article was written from the founder’s direct engineering experience and edited with AI assistance. All technical claims were reviewed before publication.
Top comments (0)