# Math Decides. AI Explains. Inside the Engine Room of Supastro

Inside Supastro: proprietary Vedic engines, a 4,892-passage classical library, multi-model AI, voice with sentiment and intent, and 1,770 accuracy checks...

Author: Sandesh · Published: 2026-10-01 · Updated: 2026-10-04 · Section: Founder Notes · Category: Behind the Build · Reading time: 11 min · Publisher: Supastro (Quantum Karma), Bangalore, India

There's a fundamental difference between an AI wrapper and an AI-native product. From the outside they look identical: a text box, a fast reply, confident sentences. You only see the difference when you open them up.

Most "AI apps" are a conventional product with a model call attached. An Ask AI button. A generated summary. A chatbot in the corner. Take the AI out and the product carries on, because the AI was never holding anything up.

Supastro is built the other way round. AI sits between your astrological data, the classical rules, your context, your intent and everything you see on screen. Take it out and there is no Oracle, no Council, no agents and no voice journal left to use.

That doesn't mean the AI does everything. It means the AI has a precisely defined job, and that job is what this note is about.

> **The short version**
>
> A wrapper adds AI to a product. Supastro is designed around it. Deterministic Vedic engines compute your chart. An AI layer of retrieval, reranking, tool-using agents, verification and memory turns that chart into answers, plans and conversations for one person.

## Four levels of AI integration

Every product that uses AI sits somewhere on this ladder. The fastest way to place one is a single question: what breaks if you remove the AI?

| Level | AI's role |
| --- | --- |
| 1. AI wrapper | Product → AI feature. An add-on, usually one call to a language model: an Ask AI button, generated copy, a summary, a chatbot. Remove it and almost nothing changes. |
| 2. AI-enabled | Product + AI. Improves parts of a conventional product: recommendations, search, classification, automation, personalization. Remove it and the product gets worse, but still works. |
| 3. AI-native | Product × AI. A core primitive the product is designed around: interpretation, reasoning, natural-language interaction, experiences that adapt. Remove it and the core experience stops making sense. |
| 4. GenAI-native | Data + models + context + generation + personalization. Takes part in creating the experience itself. It reads structured and unstructured context, reasons across it, generates, converses, adapts and coordinates several AI capabilities at once. Remove it and the experience cannot exist in the same form. |

**Level four is where we are building Supastro**, and not because generative AI is fashionable. A good Vedic reading has always been generative. A jyotishi takes fixed mathematics and classical rules and composes an answer for one person, one question and one moment in their life. That is exactly the work generative AI does well, as long as it is handed the right facts.

## Three ways to build an astrology app

| Build | The pipeline |
| --- | --- |
| Traditional app | Birth data → calculations → pre-written interpretations → you |
| AI-enhanced app | Birth data → calculations → an AI writes an interpretation → you |
| Supastro | Birth data + Vedic computation + classical knowledge + your context + your intent → retrieval, reasoning and verification → a personal answer → an experience that adapts to you |

In the first two, everyone with Saturn in the 7th house reads roughly the same paragraph. In the third, the question you asked, the dasha you're running this month, your planetary strengths and the life events you've already confirmed all change the answer.

That's why AI isn't a feature of Supastro. **It's part of the substrate.** It shapes how your chart is interpreted, how explanations are written, how the product adapts to you and how you talk to your own data, by text or by voice. It moves astrology software from one sentence to another.

> **The old model**
>
> "Here is your chart. Here are some predefined interpretations."

> **The Supastro model**
>
> "Here is your astrological data. Now let's understand what it means for your question, your circumstances and where you are in your life."

> The difference isn't that Supastro uses AI. It's what the AI is allowed to do inside the product.

## What the AI is allowed to do

Here is the AI layer as it runs in production today, job by job. The numbers come straight from our code and production database.

| Job | How it works |
| --- | --- |
| Understand | Every Oracle question is sorted by intent and screened before anything expensive runs. 4 intent classes set depth, temperature and token budget, and a gatekeeper model checks for prompt injection and off-topic requests. |
| Retrieve | Meaning search and exact-term search run in parallel over 4,892 classical passages: 1,024-dimension embeddings in pgvector plus Postgres full-text search, merged with Reciprocal Rank Fusion (k = 60). |
| Rerank | A model re-reads a shortlist of up to 24 passages and scores each one 0 to 10 against your exact question. Scores are weighted by authority (canonical 1.30, classical 1.15, curated 1.00, experimental 0.82), near-duplicates are dropped, and a source supplies at most 2 passages when alternatives exist. |
| Ground | Computed evidence, not model memory, goes into the prompt: your live dasha to the day, scored timing windows, Ashtakavarga strength, corroboration in the matching varga (D9, D10, D7, D24) and your confirmed life events. |
| Plan and act | Agents turn a prediction into a plan: dated windows linked to evidence, action steps and checklists. 10 agent workflows, a planner over 5 fixed tools (chart, live dasha, timing windows, transits, scripture search), 13 app-owned skills such as calendar holds, reminders and .ics exports, and handoffs between agents. |
| Deliberate | For big decisions, the Council argues it out before you see a word. 3 specialist agents work from the same facts, a sceptical verifier checks them, and a chief agent writes one plan that keeps dissent visible. Confidence is weighted by how well each timing engine matched your past. If the verifier fails, the Council fails closed. |
| Verify | Oracle answers are held on the server until a deterministic gate clears them. It checks completeness, the order of your life events, every chart figure and exact timing windows, with one corrective rewrite on the same model. |
| Remember | Context carries across sessions. Memory keeps patterns, not transcripts: your topics and emotional trend, confirmed life events and earlier predictions, so last month's answer isn't contradicted. |
| Listen and speak | Ask by voice and hear the answer back. Speech-to-text with acoustic sentiment, then intent, emotion and urgency read from your words, so the reply's tone fits the moment. |
| Speak first | A daily Sentinel watches your chart and reaches out when something real changes: dasha boundaries, Sade Sati, slow-planet transits. At most 3 alerts per daily run. The math picks the moment and the AI writes the note. |
| Route | Each job goes to the model best suited to it: fast models for classification and screening, larger models for readings, with automatic failover between providers. One Oracle answer stays on one model from the first word to the last. |

## And what it is never allowed to do

Half of being AI-native is deciding where AI doesn't belong. In Supastro, the AI may not:

- **Calculate your chart.** Positions, dashas, divisional charts and strengths come from deterministic engines, cached under a SHA-256 fingerprint of your birth details. Same birth, same chart, every time.
- **Interpret without sources.** Oracle answers have to cite the classical passages they were given, and Nadi and other speculative texts are kept out of Parashari readings entirely.
- **Invent a capability.** Agents choose tools and actions from fixed, app-owned lists. A tool that isn't on the list ends the plan, and a date with no computed evidence behind it can never become an action.
- **Get your dasha wrong.** An answer that names a period you aren't running, or a dasha date that doesn't exist, triggers a rewrite with the computed dates before you see it.
- **Answer a crisis.** Messages that signal self-harm get a fixed, human-written response with helpline numbers, before any model sees them.

## Anyone can call an LLM API

That part is a commodity, and it gets cheaper every quarter. What can't be copied quickly is how deeply the AI is wired into the data, the workflows, the domain knowledge and the experience around it.

For Supastro, that means an intelligence layer built specifically for Vedic astrology, rather than astrology sent as a thin prompt to a general-purpose chatbot. In practice:

- 69,000 lines of Vedic computation engines that the AI reads from and is checked against
- 4,892 classical passages from 31 sources, ranked by scriptural authority
- 10 agent workflows, a 13-skill action catalog and a multi-agent Council
- 1,770 automated accuracy checks that run before any release ships

None of that comes from adding a chatbot to an astrology app.

## The philosophy in five lines

| Approach | Where it starts |
| --- | --- |
| AI wrapper | "Let's add AI to our product." |
| AI-enabled | "Let's use AI to improve our product." |
| AI-native | "Let's design our product around what AI makes possible." |
| GenAI-native | "Let's build experiences that could not exist in the same form without generative AI." |
| Supastro | Built from the ground up for AI-native Vedic astrology. |

> **What Supastro is**
>
> Supastro is an AI-native AstroTech platform building computational infrastructure for Vedic astrology. It combines a Vedic computation engine, structured classical knowledge, temporal reasoning and generative AI to create personalized decision-intelligence agents for career, relationships, wellbeing, finance and life.
>
> **Decision intelligence, powered by Vedic computation.**

That's the claim. The rest of this note is the evidence: the engines, the library, the models, the agents and the checks, with real numbers from our codebase. It starts with an audit that showed us exactly what happens when an AI is given the wrong job.

---

## 347 planets nobody had flagged

Earlier this year we audited 250 real birth charts from our own cache. Nine planets each, 2,250 rows. The question was boring on purpose: how many of these planets are exalted or debilitated?

The answer was 347. The raw chart data had flagged **none of them**.

Not a few. None. A reading built on that data alone would treat an exalted Jupiter as an ordinary Jupiter, and a debilitated Saturn as just Saturn. Hand those rows to a language model and it fills the gap the only way it knows: fluently, confidently, and wrong.

That audit explains why [Supastro](https://supastro.com) looks the way it does. It's why we write engines before we write prompts. And it's why I can tell you exactly what sits under the hood, with numbers.

> **The one rule**
>
> Math decides. AI explains. The engines compute your chart and the AI explains what they found. It never makes up the astrology.

## The rule above my desk

Every decision at Supastro answers to that rule. Planetary strength, the dasha you're running, the yogas that are actually switched on in your chart: deterministic engines compute all of it. Give them the same birth details and they give the same answer, every time.

Only then does a frontier model get to speak. And it speaks about facts it was handed, not facts it imagined.

> Language models are brilliant writers and unreliable astronomers. So we never ask them to be astronomers.

## What's under the hood, in numbers

People ask how big Supastro really is. Here's the honest count, taken from our codebase and production database on 1 October 2026.

| What | Count |
| --- | --- |
| Source code | 333,000+ lines, including ~270,000 lines of TypeScript |
| Vedic computation engines | 69,000 lines across 147 modules |
| API endpoints | 269 endpoints serving 333 operations |
| Screens | 131 pages, 216 interface components |
| Database | 97 tables and views, 29 database functions |
| Automated accuracy checks | 1,770 across 26 test suites, on every release |
| Classical library | 4,892 passages from 31 sources |
| Classical works cited in the engines | About 18. BPHS alone appears around 390 times |
| AI reading features | 25+, plus 10 agent workflows |
| Engineering pace | 1,000+ commits in under eight months |

Lines of code mean little on their own. What matters is where they sit. Sixty-nine thousand of ours do one job: Vedic mathematics.

---

## The engines we refused to outsource

Raw planetary positions come from a professional ephemeris partner. Transits and on-the-spot charts run on our own astronomy layer, built on an open-source engine. Every Vedic judgment on top of those positions, the work a classical astrologer does by hand, is ours.

| Engine | What it computes | Why we built it |
| --- | --- | --- |
| Shadbala | Six-fold planetary strength in virupas, per BPHS and B.V. Raman | Strength decides which placements speak loudest. Ours is deterministic and costs nothing per call |
| Vimshottari Dasha | Maha, antar and pratyantar periods, per BPHS Ch.46 | Computed in microseconds, and re-verified in production every hour |
| Divisional charts | D1 to D150, including D81, D108, D144 and D150 | The higher vargas are where fine detail lives, so we compute them ourselves |
| Dignity | Exaltation, debilitation and combustion | The raw data carried none. Our audit found 347 dignities, all unflagged |
| Nakshatra | Canonical names, naming syllables, Gandamoola | 47% of charts were missing their naming syllable before this engine |
| Yogas and doshas | Formed versus activated, with classical cancellations | A yoga that exists but never activates shouldn't be sold to you as destiny |
| Ashtakavarga | BAV, SAV, reductions and Pindas, per BPHS Ch.67–71 | These numbers are what we check the AI's answers against |
| Birth time rectification | About 40 modules, with a blind benchmark | An uncertain birth time quietly corrupts everything downstream |

The engines cite their sources. BPHS, Phaladeepika, Saravali, the Jaimini Sutras, Jataka Parijata and Prashna Marga, about 18 classical works in all. When a yoga appears in your reading, it carries the source it came from and the chart factors that triggered it.

> **Why build instead of buy**
>
> Each engine exists because something broke without it. Dignity, because the data carried none. Nakshatra resolution, because almost half of all charts were missing a syllable. Rectification, because a birth time that's off by minutes can move your Lagna in the divisional charts.

---

## A library the AI has to read first

Engines handle the math. Interpretation needs the texts.

So we built a library: 4,892 passages from 31 sources. Around 4,100 of them come from eight classical treatises: BPHS, Brihat Jataka, Jataka Parijata, Phaladeepika, Prasna Marga, Uttara Kalamrita, Laghu Parashari and Deva Keralam. Each passage is stored as a 1,024-dimension vector inside our own Postgres database.

When you ask a question, two searches run at once. One searches by meaning, the other by the exact term, and the results are merged. Jyotish questions often turn on a single word like Gandanta or Arishta, and a search by meaning alone tends to blur those. Results are then ranked by scriptural authority, so a line from BPHS outranks a modern commentary, and near-duplicates are dropped before anything reaches the model.

_Figure: How a Supastro answer is made: the engines compute, the classical library grounds, and the integrity gate checks every answer before delivery._

---

## Choosing the models

We don't marry one model. Each job goes to the model that does it best, and if a provider has a bad minute, the request fails over to the next one automatically. A single Oracle answer, though, stays on one model from the first word to the last, so it never changes voice halfway through.

Small, fast models do the quick work. They sort your question by intent and screen for prompt-injection attempts before anything expensive runs. Larger models write the readings. The premium report has its own model, chosen for long, structured writing.

| Provider | Model | Where it works | Status |
| --- | --- | --- | --- |
| Google | Gemini 3.1 Pro | Feature specific | Live |
| Google | Gemini 3.5 Flash-Lite | Feature specific | Live |
| Google | Gemini 3.8 Flash | Feature specific | In internal testing |
| Google | Gemini Embedding | 1,024-dimension vectors for the classical library | Live |
| OpenAI | GPT-5.6 Sol | Feature specific | Live |
| OpenAI | GPT-6 Astra | Feature specific | In internal testing |
| OpenAI | GPT-6.1 Sol | Feature specific | In internal testing |
| Anthropic | Claude Sonnet 4.6 | Feature specific | Live |
| Anthropic | Claude Sonnet 5.5 | Feature specific | In internal testing |
| Anthropic | Claude Opus 5.5 | Feature specific | In internal testing |
| Anthropic | Claude Fable 5.1 | Feature specific | In internal testing |
| Deepgram | Nova-2 and Aura-2 | Speech-to-text with sentiment and intent; spoken replies | Live |

> **A note on the model list**
>
> A few of these models are in internal testing and will roll out soon, once they pass our accuracy checks. Providers update their models often, so this lineup will keep changing.
>
> _“Supastro dynamically routes each astrological question to the appropriate reasoning model based on complexity, evidence requirements and response depth”_

Voice gets the same care. Deepgram turns speech into text, and on voice journal entries it also reads **sentiment and intent**, so the reply knows whether you're anxious about a deadline or just curious about Saturn. The score never appears on your screen. It only shapes how the answer is written. When Supastro talks back, Deepgram's Aura voice does the speaking.

## Agents that have to show their work

Supastro runs ten agent workflows: Council, Daily Briefing, Decision Room, Karmic Life OS, Sadhana, Saturn Companion, Sentinel, Big Day, Destiny Timing and Muhurta. The agent runtime plans each step and picks from a fixed set of five tools: chart summary, live dasha, timing windows, current transits and scripture search. Then it has to answer using only what those tools returned.

Then comes the part I'm proudest of. Before any Oracle answer reaches you, an integrity gate checks it against your computed chart across seven evidence layers: natal placements, divisional charts, dashas, strength, transits, yogas and special points. A wrong Ashtakavarga number, a dasha date that doesn't exist, a claim with nothing behind it: the gate catches it, and the answer is rewritten once with the correction. If it still isn't clean, you don't get it.

> **Fair by design**
>
> Credits are reserved when you ask. If the answer fails a check or times out, they come back to you automatically.

---

## Why this stack

Every tool in the stack had to earn its place. Here's the reasoning, layer by layer.

_Figure: The Supastro stack: six layers, from the interface down to operations._

| Layer | What we use | Why we chose it |
| --- | --- | --- |
| Interface | Next.js 16, React 19, Tailwind CSS 4 | Server-rendered pages load fast and stay readable to search engines. One framework serves both the screens and the API |
| Language | TypeScript, end to end | A chart is hundreds of typed facts. A misspelled field should fail the build, not your reading |
| Request layer | Node.js proxy | Refreshes your session securely on every request, and serves clean Markdown to AI answer engines |
| API | Serverless functions on Vercel | Oracle answers stream live, and capacity grows with demand instead of sitting idle |
| Database | Supabase Postgres with row-level security and pgvector | Accounts, the credit ledger and the library's vectors live in one place, and each person can only ever read their own rows |
| Payments | Razorpay in India, Dodo Payments internationally | Local payment methods in each region, and every payment is fulfilled exactly once |
| Operations | Sentry, PostHog and 14 scheduled jobs | Errors surface in minutes, and background jobs reconcile payments, send reminders and self-test the engines |

None of it ships on hope. 1,770 automated checks across 26 test suites run on every code push and every production build. If one fails, the release stops. In production, a golden chart is recomputed every hour to confirm the dasha math hasn't drifted.

---

## Where this goes next

I want Supastro to be deeptech. Not as a slogan, but as an accurate description of the work. That means machine learning, computer vision and speech research built on top of the engines, never in place of them.

| Track | What we're exploring | Status |
| --- | --- | --- |
| Machine learning | Timing models calibrated on consented, anonymised life-event feedback | Research |
| Computer vision | Our own palm-line detection, and reading Sanskrit and regional-language manuscripts to grow the library | Research |
| Speech | Voice in Hindi and regional languages, with tone-aware conversation | Planned |
| Knowledge graph | Classical rules as a graph the engines can query and cite | Research |
| Astronomy | A fully in-house ephemeris | Planned |
| Rectification | Learning from verified birth times, scored in a blind benchmark | Benchmark built, research ongoing |

Let me be straight about timing: most of this is a long way off. Research done properly takes years, and I'd rather ship one model that holds up in a blind test than five that demo well.

Right now the work is less glamorous and more important: making today's product flawless, neat and deeply personal to each chart. Every bug fixed, every edge case closed, every engine tightened.

---

## The point of all this

Astrology has a credibility problem, and a lot of it is earned. Vague readings. Invented yogas. Apps that are autocomplete in a saffron robe.

[Supastro](https://supastro.com) is my answer to that. Real mathematics. Classical sources you can trace. Frontier AI that has to show its work. And a system that checks itself 1,770 times before it's allowed to change.

Build your chart and ask the hardest question you have. Then check the answer.

Sandesh

### Keep reading

- [How Supastro Is Built](https://supastro.com/founder-notes/how-supastro-is-built) — The five layers, from astronomy to agents
- [The RAG Stack Behind Every Prediction](https://supastro.com/founder-notes/the-rag-stack-behind-every-prediction) — How the classical library is searched
- [Why LLMs Hallucinate Astrology](https://supastro.com/founder-notes/why-llms-hallucinate-astrology) — And how the integrity gate stops it
- [Why Planetary Strength Matters](https://supastro.com/founder-notes/why-planetary-strength-matters-in-ai-astrology) — Shadbala, explained

**Ask the hardest question you have** Build your chart and watch every engine work, from strength to dasha to the classical source behind each line. [Open Supastro](https://supastro.com/dashboard)

---

Canonical: https://supastro.com/founder-notes/math-decides-ai-explains-inside-supastro · Supastro is an AI astrology platform for Vedic astrology (Jyotish), built as Vedic Super Intelligence (SI), by Quantum Karma, Bangalore, India. Chart positions are computed from NASA JPL ephemeris data on the Lahiri (Chitrapaksha) ayanamsa. 40 instruments are free to use.