§

Project ABHIGYAN

A Learning Engine for K-12

An Educational Brief

By Pragnan Chakraborty [M.A. Computational Linguistics]

A Tutor for Every Child

A system built on a simple idea from learning science: you cannot teach a child well until you understand how that child currently thinks. This brief outlines the problem, and the approach we have built around it, for your review.

For educator review Focus K-12 · ages 6 to 16 Grounded in learning science

Five children sit through the same lesson. One of them was taught at the right level.

A devoted teacher cannot, in real time, hold a separate mental model of every child: who has the prerequisite, who is quietly lost, who is bored, who is carrying a misconception that will surface three chapters later. The lesson is pitched at the middle, and the rest are served by accident.

Figure 1 · One Lesson, Five Children
MASTERY LESSON IDEA 1 IDEA 2 IDEA 3 IDEA 4 IDEA 5 ★
Well matched: the lesson happened to meet this child where they were. Reaches mastery.
Missing the prerequisite: lost at the first new idea, with no time to catch it.
Already ahead: re-learns what they know, attention drifts.
Quietly disengaged: never signals confusion, slips further behind unseen.
Hidden misconception: looks fine on the surface, but a wrong mental model goes uncaught.

The lesson was the same for all five. Only one of them was actually taught.

This is not a new problem. Learning science named it forty years ago.

A child tutored one-to-one performs about two standard deviations above a child in an ordinary classroom.1

Benjamin Bloom's 1984 finding remains one of the most cited results in education: the average tutored student outperformed roughly 98% of students taught conventionally.1 The reason is not that tutors know more. It is that a tutor continuously diagnoses the learner, corrects misconceptions the moment they appear, and adjusts pace and difficulty to the individual, the conditions of mastery learning.1,2

Bloom called the search for scalable methods that match tutoring "the two-sigma problem."1 It has stayed open because a human tutor for every child has never been affordable.

2σ
Tutoring advantage
Bloom, 1984. One-to-one + mastery learning vs. conventional class.
98%
Outperformed
Share of the conventional class the average tutored child beats.
1:30
The constraint
The teacher-to-child ratio that makes true tutoring impossible at scale.

The question this brief addresses is Bloom's question: can the diagnostic attention of a tutor be offered to every child, without replacing the teacher who knows them?

The thread that runs through tutoring is diagnosis. So we begin there.

Formative assessment,4,5 Vygotsky's zone of proximal development,3 mastery learning2: these ideas share one root. Good teaching meets a child just beyond what they can already do, and is guided by a clear read of where that edge is. The shift below is the whole approach.

Teaching the Class

Deliver the syllabus

  • One pace, pitched at the middle
  • Marks the answer right or wrong
  • Misconceptions surface in the exam, too late
For the child: covered, but not necessarily understood
Teaching the Child

Build a model of understanding

  • Pace and difficulty set per child
  • Reads why an answer was wrong
  • Catches the misconception as it forms
For the child: met where they are, moved forward from there
If diagnosis is the principle, something has to do the diagnosing, continuously.

At the centre is what we call a cognitive profile: a structured, evolving picture of a child's grasp of each idea, their pace, and their habits of reasoning. It mirrors the mental model a good tutor builds of a student, and it is updated through the same loop a tutor runs intuitively.

TEACH
Present
An idea, pitched to the child's current level
→
CHECK
Diagnose
Read why an answer was right or wrong, not just the score
→
ADAPT
Adjust
Update the profile, then re-pitch the next idea
↺

This is the formative-assessment cycle,4,5 run patiently for one child at a time, at a scale no single teacher's day allows.

What does that living model actually look like? Here is one child's.

A profile for a Class 8 student in Mathematics. Hover a weak area to see the kind of evidence the engine holds: not a mark, but a read of the underlying misconception,6 expressed in terms a teacher would recognise.

Cognitive Profile · Class 8 · MathematicsLIVE
Integers
84
Fractions
41
Adds numerators and denominators directly (1/2 + 1/3 = 2/5). A part-whole misconception, not careless error. Needs the idea of a common unit, not more drill.
Geometry
67
Word problems
38
Computes correctly once the equation is set up, but cannot translate the sentence into the equation. The gap is comprehension and modelling, not arithmetic.

The same wrong answer can come from very different causes. Teaching that ignores the cause repeats the lesson. Teaching that names it can fix it.

A diagnosis points to a response. Ours is a small, structured classroom.

Before a session begins, three planning steps run against the child's profile to design a short, Socratic class: what to cover, which peer voices to include, and where to check understanding. Press play to watch the planning, or click a step to read what it does.

Figure 2 · Planning a Session for One Child
PLANNER
READS THE PROFILE
01
Lesson Designer
Chooses the idea to focus on and the order to build it, starting from what the child already holds.
02
Peer Voices
Creates classmate characters who think differently, so the child sees more than one way to reason.
03
Check Designer
Places gentle questions through the session to confirm understanding before moving on.
Press Play or Step to begin
Planning Log · One Session
Waiting for planning to begin…
The child sees a friendly class. This planning stays behind the scenes.
Much of a child's real thinking happens on paper, not on a screen.

A photograph of a handwritten solution, read and mapped back to skills.

Much of what reveals a child's thinking lives in their working: the step where a sign flipped, the line where a method was misapplied. A child photographs their handwritten answer; the engine assesses it and maps the feedback to specific ideas in the profile, so the next session can respond to it.

  • Reads the steps, not just the result, so partial understanding is recognised and credited.
  • Locates where reasoning broke, turning a wrong answer into a specific, teachable moment.
  • Keeps pencil-and-paper central, rather than pushing every task onto a keyboard.

This matters most for younger learners, where writing by hand is part of how the thinking itself is formed.

None of this is meant to stand in for the teacher. It is meant to reach them.

A teacher cannot maintain thirty live cognitive profiles in their head, but they are irreplaceable at everything that follows from one: knowing the child, motivating them, deciding what truly matters. The engine is designed to give the teacher that diagnostic layer, and to step back.

01
Visibility into every child

The teacher sees, at a glance, who is stuck on what, and why, across the whole class.

02
The grunt-work, handled

Continuous diagnosis and routine practice are carried by the engine, freeing teacher time for teaching.

03
Judgement stays human

The engine surfaces evidence and suggestions. What to do with a particular child remains the teacher's call.

04
A shared language

Profiles are expressed in pedagogical terms, so they inform parent and teacher conversations, not replace them.

Any system that sits with a child must earn a different kind of trust.

We treat this as the first requirement, not an afterthought. The design principles below are the ones we would want any system working with our own children to meet, and the ones we most want your scrutiny on.

Oversight
Teacher and parent in the loop. The engine never sits alone with a child as the only adult presence. Activity is visible to the adults responsible.
Transparency
No black box. Every judgement about a child is expressed as readable evidence a teacher can question and override.
Age-appropriate
Content and tone bounded by age. Material, language, and interaction are scoped to the child's stage, with clear limits.
Data care
Minimal, protected, India-resident. A child's data is collected sparingly, held securely, and governed by a clear consent and residency roadmap.

The aim is never to isolate a child with a machine, but to give the adults around them a clearer view of how to help.

One last point on where this sits in the wider work.

The same cognitive engine, the diagnosis loop, the profile, the Socratic session, also powers a platform for adult competitive-exam aspirants. The underlying model of a learner is general. What changes for K-12 is everything that should change for a child:

  • Tone and pacing built for younger learners, warm rather than clinical.
  • Curriculum alignment to school boards and grade-level expectations.
  • Stronger safeguards, with teacher and parent oversight built in from the start.

K-12 does not depend on the rest to work. It is a complete system in its own right, sharing a proven core rather than borrowing an afterthought.

Which brings us to the reason this brief is in your hands.

The engineering is ours to get right. The pedagogy, we would like to get right with you.

Question One
Does the diagnosis-first approach hold up against how children actually learn at each stage?
Question Two
Where might a system like this mislead a teacher, or a child, and how should it guard against that?
Question Three
What would make this genuinely useful in a real classroom, rather than one more screen?

Thank you for reading. Your judgement on these questions is exactly the input this work needs next.

Appendix

Bibliography

1
Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16.
2
Bloom, B. S. (1968). Learning for mastery. Evaluation Comment, 1(2), 1–12. University of California, Los Angeles (CSEIP).
3
Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press. (M. Cole, V. John-Steiner, S. Scribner, & E. Souberman, Eds.)
4
Black, P., & Wiliam, D. (1998). Inside the black box: Raising standards through classroom assessment. Phi Delta Kappan, 80(2), 139–148.
5
Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education: Principles, Policy & Practice, 5(1), 7–74.
6
Ni, Y., & Zhou, Y.-D. (2005). Teaching and learning fraction and rational numbers: The origins and implications of whole number bias. Educational Psychologist, 40(1), 27–52.

These works are cited as the established ideas this approach is built upon. They describe the learning science; they are not claims of measured outcomes for this system, which is precisely what we hope to study with educators next.