How AI Astrology Actually Works: Ephemeris, Rule Engine and Language Model

"AI astrology" can mean two very different things. In one, a language model is asked about your life and writes an answer. In the other, your chart is calculated by astronomical software, classical rules are tested against it by ordinary code, and a language model is brought in only at the end — to explain what was found, under checks. Both produce fluent paragraphs. Only the second produces a reading that can be checked.

This post walks through how the second kind works, using ProxaAI's own pipeline as the example, because we can describe it exactly. The full technical version is on the methodology page.

Layer 1: Astronomy — the Ephemeris

Everything starts with a calculation that involves no AI at all. An ephemeris is a model of the solar system that gives the position of each planet for any moment. ProxaAI uses the Swiss Ephemeris, the engine behind most professional astrological software.

From your date, time and place of birth it computes, for all nine grahas:

• the sidereal sign and degree, using the Lahiri (Chitrapaksha) ayanamsa;
• the nakshatra and pada, retrograde state, dignity and combustion;
• the lagna and the houses (whole-sign, the convention the classical rules were written for);
• Shadbala, Ashtakavarga and the divisional charts, including the Navamsha;
• your Vimshottari dasha timeline, and a calendar of the slow planets for the next three years — when Jupiter, Saturn, Rahu and Ketu change sign, and when Sade Sati begins and ends.

Because a recorded birth time is rarely exact, the chart is also recomputed five minutes either side, and anything that changes inside that margin is marked conditional. This layer is deterministic: the same birth details give the same chart every time.

Layer 2: Rules — the Deterministic Engine

The second layer also involves no AI. Classical astrology is, to a large extent, a body of conditions: Gaj Kesari Yoga forms when Jupiter is in a kendra from the Moon; a Raj Yoga needs a link between the lord of a kendra and the lord of a trikona; a dosha has named cancellations. Conditions like these can be written as rules a program evaluates.

ProxaAI keeps a knowledge base of such rules. The Parashari set — 44 yogas and 10 doshas — is encoded as executable rules, each citing its source (most often Brihat Parashara Hora Shastra and Phaladeepika); Jaimini, Lal Kitab and Nadi charts add their own rule sets. A rule engine walks every rule against your chart and returns exactly the ones that form, with the classical cancellations applied, and grades each yoga by the planets that make it.

Timing works the same way. For a question about career, marriage, money, home or children, every antardasha of the next eight years is ranked by how its two lords connect to the house in question, with the classical reason for each link. Doshas that appear in no classical text — Kaal Sarp, Pitra and five others — are still computed, because people ask about them, but they are labelled modern.

This is the step most "AI astrology" skips. When a language model names your yogas, it is recalling what usually gets said. When a rule engine names them, it has tested them.

Layer 3: Language — Where the AI Comes In

Only now does a language model get involved, and its job is narrower than people assume: interpretation and language.

Your question goes to a planner, which decides which specialists it actually needs. The four core specialists are a Vedic astrologer, a numerologist, a Prashna (horary) reader and a transit-and-dasha specialist; a Western astrologer, a BaZi reader, a palmist and a face reader join when your kundli follows those traditions or you have added palm and face photos. Each specialist receives the computed facts for its own system — the numerologist gets the numbers of your birth date and name, never your birth chart — and a filtered slice of the rules that apply.

The specialists are instructed not to contradict the computed chart, not to invent yogas, doshas or dasha lords beyond those supplied, and not to cite a classical rule they were not given. On a yes-or-no question, the Prashna specialist's own judgement comes from the rule engine and the model may not reverse it; the final verdict combines it with the other specialists' verdicts by fixed weights.

Layer 4: Checking — Why the Answer Is Validated

Instructions are not enough, so the specialists' notes are checked before the answer is written, each against what was computed for it. The checks differ by specialist:

• in the Vedic astrologer's notes, a named yoga that does not exist in the knowledge base is flagged as fabricated, one that exists but did not form for your chart is flagged too, and signs and dignities are checked against the computed positions;
• in the Vedic astrologer's and the transit specialist's notes, a dated dasha period your timeline does not contain is caught, and the numerology, Western and BaZi notes may not name a dasha at all;
• a timing window from the astrology or transit specialist that matches none of your computed dasha periods or transit spans is removed.

The default response to a failed check is removal from the notes, not rewording. A supervisor then writes one answer from the checked notes. Where the specialists disagree, the weighting is fixed in code rather than decided by the model; methods that read the same data count once; and the answer is given in words — leans yes, uncertain or leans no — with a star rating for how strong and how independent the agreeing evidence is. The answer is also checked as it is written: its verdict line must match the computed verdict, and percentages, probabilities and odds are taken out. We do not give a percentage, because no method here has been calibrated against outcomes well enough to earn one.

One Question, Followed Through

Take a real-world question: "Is the next year a good time to change jobs?"

1. Layer 1 has already computed the chart, the 10th house and its lord, the dasha periods running through the next year and Saturn's and Jupiter's transits.
2. Layer 2 ranks the coming antardashas by how their lords connect to the 10th house, and judges each transit from the Moon sign by Phaladeepika's rules.
3. Layer 3: the planner picks the specialists the question needs — for a timing question like this, typically the Vedic astrologer and the transit-and-dasha specialist. Each writes its reading from those facts.
4. Layer 4 checks the specialists' notes. A window that does not line up with a computed period is removed, and so is a yoga the chart does not have from the Vedic astrologer's notes.
5. The supervisor gives the most likely window with the dasha period behind it — or says plainly that no clear window appears — and rates its confidence.

A Chatbot-Only Approach, Compared

The ChatGPT kundli test shows how to tell the two apart in five minutes, on any tool.

What AI Astrology Cannot Do

A correct calculation is not a correct future. Astrology describes tendencies and timing; read every reading, ours included, as guidance rather than a promise. The order in which the rule engine weighs its reasons is our own and has not been calibrated against known event dates. Readings are built not to predict death or lifespan or diagnose an illness, and they are not a substitute for medical, legal or financial advice.

What we stand behind is narrower and real: the astronomy is computed to a published standard, the classical rules are evaluated by code and cite their sources — each specialist works from cited rules in its own tradition, and where texts disagree we follow one stated school — and the reasoning is shown, so you, or any astrologer you trust, can check it.

See It on Your Own Chart

Start with the free kundli: Layer 1, with no sign-up. To see all four layers at work on a real question, ask the AI astrologer — your first question is free. And if you would like the whole chart read at once, the Complete Kundli Report runs the same pipeline across twelve traditions — Vedic Parashari, KP, Jaimini, Lal Kitab, Nadi, Western, BaZi, numerology, Nine Star Ki, Mahabote, Mayan and Tibetan — into a 20-page PDF for ₹499.

For a worked example of a rule being tested rather than recalled, read how Raj Yoga forms and why many charts have none.