WITH CLAUDE logo WITH CLAUDE Search

KAKSHA

A practice simulator for trainee teachers: Claude plays five Grade 3 pupils with hidden learning levels and misconceptions, then a coach reviews what the teacher missed.

The problem

Teacher training in India runs at enormous scale and doesn't change what happens in classrooms. NISHTHA has reached 4.2 million teacher-instances; independent evaluations by Oxford RECOUP and Azim Premji University found no significant gain in Grade 5 maths outcomes. The system optimises for training delivered, not practice changed — because there is nowhere safe to practise. D.El.Ed and B.Ed trainees get a few unobserved weeks in a real school. DIETs, the institutions meant to supply guided practice, carry 30–50% faculty vacancies. So new teachers learn to read a classroom on real children, and the cost of every early mistake is paid by a child who can't yet read.

How it works

KAKSHA is a flight simulator for teachers. Claude plays five Grade 3 children in an Indian government school, each with a hidden learning level and a hidden misconception the teacher cannot see. Reading levels follow the ASER 2024 bands; the spread across the five is the documented ~7 grade-equivalents found inside a single grade; the maths errors are the classic documented subtraction bugs.

You teach six turns, blind. Two model calls drive it. The first runs the classroom: five personas responding live, wrong answers derived deterministically from each child's rule rather than randomly — Priya returns 35 for 52−27 every single run — so the misconception is consistent and therefore diagnosable. A bug stays hidden until a question would expose it: Imran looks fine until you set a sum containing a zero. Engagement decays when a child is ignored.

The second call is the coach. It receives the transcript and the five hidden profiles, then scores attention equity, question quality and misconceptions caught, and names the moments you missed. The lesson replays with every child's actual thoughts revealed — the gap between what you saw and what was happening is the learning event.

It solves the practice problem without the blockers that kill real-classroom observation tools: no recording, no consent, no child data, no facilitator per trainee.

How Claude was used

In the team’s own words, from their submission.

Claude plays five Grade 3 children in an Indian government school, each with a hidden learning level and a hidden misconception the teacher cannot see.

Team

Individual builders are not listed yet — names are published only with their permission.

← Back to the Project Directory