Problem
The one domain where a plausible answer is worthless
Vedic astrology is, underneath the interpretation, an exact astronomical calculation. Planetary positions, house divisions, the ayanamsa correction, the nested period system — these are arithmetic with correct answers. The interpretive layer above them is where the tradition, the judgement and the ambiguity live.
Which makes it the worst possible candidate for a language model handling the whole request end to end. Ask a model for a chart and it produces something that looks exactly like a chart: correctly formatted, confidently stated, and numerically wrong. There is no user-visible difference between a computed position and a hallucinated one — which is precisely why the two have to be produced by different machinery.
Architecture
A calculation engine with a language layer bolted above it, never into it
The system is built in two strictly separated halves. The lower half is deterministic: it takes a time, a place and a coordinate system and computes the chart — positions, divisional charts, period sequences — as arithmetic. It has no model calls in it at all, and its output is fully reproducible for the same inputs.
The upper half is where language belongs: explaining what a placement means, answering a question in the user's own words, surfacing what classical sources say about a combination. It is only ever allowed to read the computed chart. It cannot recompute, adjust, round or infer any figure in it.
Split the system on the question “does this have a correct answer?” Everything that does is code. Everything that doesn't is where a model earns its place.
That split also decides the caching and cost story. Computation is deterministic and therefore cacheable forever; generation is variable and priced per call. Keeping them apart means the expensive half runs only on the interpretive question, never on regenerating a chart the system already knows.
AI components
Language over structure, never language instead of structure
- Natural-language querying — a user's question is classified and resolved into a structured lookup against the computed chart, rather than answered from the model's own recollection of the subject.
- Retrieval over source material — interpretive text is grounded in a corpus of 9,258 classical verses, so an explanation can be traced back to a source rather than invented in the moment. This is the difference between a chart reading and a plausible-sounding paragraph.
- Structured extraction — birth details supplied as free text (a date written six different ways, a place name rather than coordinates) are normalised into typed inputs before the calculation engine sees them.
- Generation with a fixed factual frame — every generated passage is produced against computed values passed in as context. The model writes the prose; the numbers arrive already decided.
- Deliberately not AI — positions, divisional charts, period calculations and date arithmetic. All of it is code, all of it reproducible, none of it delegated.
What it runs today
A working research tool, not a novelty
VedicPupil is live at vedicpupil.in, serving chart generation, a verse library, reports and a chat surface where a user can question their own chart in plain language — the deterministic engine answering the exact questions, the retrieval and language layer handling explanation on top of it.
The architecture has held through model changes underneath it. Because generation only ever consumes computed output, swapping or upgrading the model changes how an explanation reads and never changes what the chart says.
Outcome
The boundary is the product
The useful lesson from this system generalises well past astrology. Every domain has a layer with correct answers and a layer without, and most disappointing AI products are built by a team that never drew the line between them — so the model gets handed the arithmetic along with the prose, and quietly gets some of it wrong.
Finding that line, and then enforcing it in the architecture rather than in a prompt, is most of what we do on an engagement. VedicPupil is the cleanest illustration of it we have, because in this domain a wrong number and a right number look identical on the page.