The companion
who starts at zero.

Shunya — zero. Saarthi — the one who drives the chariot, the guide who stays beside you. In the Mahabharata, the saarthi is not the warrior. The saarthi sees clearly, stays steady, and asks the question that helps the warrior find their own answer. ShunyaSaarthi positions itself not as a content delivery machine but as a companion that starts where the learner actually is — at zero — and goes wherever the learner goes.

That is the product's implicit contract with its users, and the rest of the architecture either honors it or violates it. So far, most of the architectural choices honor it well.

It also locates the product within a specific tradition: Indian secondary education, NCERT textbooks, the CBSE examination structure. This is not a generic "AI tutor." It is a companion for Class 6 through Class 12 students in India, grounded in the curriculum they are actually tested on. That specificity is a strength, not a constraint.

India has roughly 250 million secondary school students. The majority attend government or low-fee private schools where a single teacher covers 40 to 60 students, where the textbook is the curriculum, and where a student who misses the intuition behind a concept in Chapter 1 will struggle silently through every chapter that follows.

NCERT textbooks are excellent — genuinely world-class in their structure, sequencing, and rigor. They are also static. A confused student cannot ask the textbook a question. The textbook cannot notice that the student keeps getting sign errors in Coulomb's Law because they never quite internalized the direction convention. It cannot slow down, try a different example, or check whether the Class 9 concept of potential energy is actually solid before building on it.

Private coaching centers and tutors fill this gap for families who can afford them. ShunyaSaarthi is building the AI layer that could fill it for those who cannot — not by replacing the teacher, but by giving every student access to something that knows the curriculum deeply, responds to what the student actually says, and refuses to let a misconception go unchallenged.

Most AI tutoring products are wrappers: a well-prompted LLM over a PDF, with perhaps some retrieval-augmented generation on top. ShunyaSaarthi is not that. It has made three architectural commitments that are non-trivial, mutually reinforcing, and not reversible once the system is at scale.

Bet One: The knowledge base is NCERT-grounded, not LLM-generated

Every concept, question, and formula in the system traces to a specific NCERT section. ncert_citation is a required field, not an optional annotation. The LLM cannot invent curriculum or drift into content the student's actual exam will not cover.

This sounds like a constraint, but it is actually what makes the system trustworthy to a parent or teacher. The answer to "where did this question come from?" is always "Chapter 2, Section 2.1." That chain of accountability is load-bearing.

Bet Two: The authority boundary — Python owns evidence, LLM owns voice

This is the most important architectural decision in the system, and it is surprisingly rare. Python owns mastery, progression, rewards, evidence accumulation, and safety. The LLM owns narration, coaching voice, and question phrasing. These domains do not leak into each other.

This is not a guardrail bolted on top of an LLM. It is an architectural choice that gives the LLM genuine creative freedom in a bounded domain. The result is that the LLM performs better — it knows exactly what it is responsible for — and the system is safer, because the decisions that matter most are never delegated to something that hallucinates.

LLM DOMAIN              PYTHON DOMAIN
─────────────────────    ───────────────────────────
scene narration      →   milestone gate validation
coaching voice       →   mastery threshold check
question phrasing    →   answer correctness
difficulty proposal  →   evidence accumulation
branch story choice  →   reward eligibility
reflection text      →   completion status
                         checkpoint persistence
                         NCERT citation validation

Bet Three: The Learning Intelligence Generator pre-authors the lesson contract

Rather than generating lesson content at runtime from a raw textbook, the LIG pipeline pre-generates structured lesson briefs — JSON contracts that contain concept explanation, worked examples, visual specifications, misconception corrections, NCERT citations, and adaptive content instructions. The lesson brief is a contract between the knowledge base and the live tutor. It separates editorial quality control (where teacher review happens, at generation time) from runtime adaptivity (where SUNO responds to what the student actually says, in the session).

This means the platform can review the quality of a lesson in a document, not across millions of runtime LLM calls. It is a fundamentally more auditable model.

SUNO (सुनो — "listen") is the student-facing tutor. It is not a chatbot. It is an orchestrated agent with an explicit session contract: a seven-stage adaptive pipeline that moves a student from initial assessment through scaffolded learning, concept building, consolidation, extension, diagnostic re-check, and generalization. Each stage has entry conditions, exit criteria, and mastery thresholds defined in code.

The seven stages are not decorative labels. They are gated transitions with evidence requirements. A student does not move from B to C because the session clock ran out; they move because Python has accumulated sufficient correct-response evidence at the right Bloom level. The LLM proposes the next interaction; Python decides whether the gate opens.

Movie Mode: the long bet

The Movie Mode experiment — currently a POC for Class 6 Mathematics Chapter 10, "The Other Side of Zero" — shows where SUNO is pointing. It is a complete agentic lesson that feels like interactive narrative. The learner is dropped into a story world (Minecraft mines, a scoreboard, a thermometer building) and their mathematical choices change what happens next. But the story is a vehicle, not the lesson. Mathematical correctness determines progression. The LLM generates narrative; Python validates every milestone transition.

Five required milestones — World Below Zero, Twins and Opposites, Number Line Battle, Debt Trap, Final Battle — are hard-gated in order. No story branch can skip a mathematical requirement. The five-milestone structure ensures that a student who completes the story has necessarily encountered every key concept in the chapter. Story engagement is not the product. It is the delivery mechanism for genuine learning evidence.

Every generated artifact in the system carries teacher_review_required: true. This is enforced at the model layer, not just at the API. It is not a disclaimer. It is a design decision about what the product is and is not claiming.

The claim this platform makes is limited and honest: here is AI-generated content that has been produced with strong curriculum grounding, deterministic evidence, and validated citations; a trained teacher should review it before any student sees it. The platform is not claiming the AI is infallible. It is claiming that AI-plus-teacher is better than teacher-alone-with-too-little-time.

A teacher who reviews 30 AI-generated assessment items and approves 28 of them has just authored a practice paper in 10 minutes that would have taken two hours. The teacher's judgment is preserved. Their production cost is dramatically reduced. The two items they rejected taught the system something. The gate is the product, not an obstacle to it.

This is also what distinguishes ShunyaSaarthi from AI products that quietly elide the human-oversight step when it becomes inconvenient. The architecture makes teacher review structurally mandatory. That choice will create friction at scale, and the product will have to build teacher UX that makes review genuinely fast rather than nominally required. But the choice is right.

The architecture is sound. The design philosophy is coherent. The tensions that follow are not criticisms — they are the real work that remains.

• T1
  POC to production. Movie Mode, the full LIG batch across six corpora, SUNO Live — these are POC and teacher-review state today. The gap between "architecturally sound" and "serving 10,000 concurrent students without degrading the evidence quality" is real and largely uncharted. The architecture is designed to scale; the operations layer is not yet.
• T2
  The teacher review bottleneck. At 10 students, the teacher gate is a quality mechanism. At 10,000, it becomes the throughput constraint on the whole system. The platform needs teacher UX that makes review fast enough to keep pace with learning demand — not just a flag the teacher sets and forgets.
• T3
  Corpus quality variance. Physics XII is deeply enriched: 14 chapters, 86 validated formulae, diagram templates, NCERT section citations at the sentence level. The newer corpora — Science IX, Mathematics VI — are at different stages. LIG output is only as good as the underlying chapter packs. The gap in corpus quality is, in practice, a gap in the education the platform can deliver.
• T4
  Multi-agent build complexity. Three agents — Claude Code, Cursor, Codex — are building simultaneously against a frozen contract. The coordination machinery (STATE.json, claim files, the preflight check) is working well. But this coordination overhead is real, and the risk of a schema drift or a lock conflict causing a regression in a live student session grows as the system grows.
• T5
  The mastery claim question. The system is rigorous about not overclaiming mastery — the authority boundary prevents the LLM from asserting completion; Python gates every milestone. But there is an inverse risk: a system that is so careful about false positives that it underserves students who genuinely understand the material and need to move faster. Calibrating the gates for both directions is the remaining hard problem.

The Class 6 work is the most interesting signal in the system right now. Not because integers are harder than electrostatics, but because of what a Class 6 entry point implies: the platform meeting a student at age 11, learning how they think about number and space, and then staying with them through all of secondary school — through the abstraction jump of Class 9, the formalism of Class 10 boards, the conceptual density of Class 12 Physics.

That longitudinal relationship is what the name promises. Not a session. Not a practice paper. A saarthi — someone who stays in the chariot for the whole journey.

The infrastructure is being built to support this. The adaptive store accumulates evidence across sessions. The mastery model carries forward. The corpus isolation means a Class 6 result does not contaminate a Class 12 session. The authority boundary means the evidence being accumulated is trustworthy — it was earned by the student, not hallucinated by the LLM.

The honest summary: ShunyaSaarthi is doing the hard things in the right order. The knowledge base before the interface. The authority boundary before the LLM creativity. The teacher gate before the student session. The POC before the scale.

Most AI education products invert this. They build the engaging interface first, the underlying learning model later, and discover that the two were never connected in a way that survives contact with a real student. ShunyaSaarthi is building the connection first. The engaging interface is coming — Movie Mode is its early shape. The foundation that will make it trustworthy when it arrives is already in place.