Raghav Jalan

Jaipur, India

All work

2 of 6

PAI Learn

A teacher that asks before it answers.

Role
Technical co-founder, engineering
Status
In development, co-founder, 2026
Platform
Installable web app
PAI Learn's website on a desktop: “One question. Different ways in.”, an interactive concept demonstration, with a note that PAI Learn is in development.
PAI Learn's website on a phone: “A new kind of learning companion. One question.”

In short

Who it’s for
A 14 to 16-year-old learning IGCSE maths, and the parent who wants to know how they’re really doing.
What was hard
“I don’t understand” can mean very different things, and most tools respond to all of them the same way.
What I decided
Understand before teaching: a rules engine chooses the kind of help, the AI writes only within those limits, and every reply is checked before a child sees it.

The person who once taught in my school is now the person I’m building with.

In 2020, Shariq Khaleel came to my school to run a workshop. I was one of the students in the room.

I didn’t know then that it was the start of something. He became my mentor, and we stayed in touch as I taught myself to build software. Over the years, the student in the audience slowly became someone who could build what he imagined.

In 2026, we started PAI Learn together. Shariq leads education and product. I lead the engineering: the AI architecture and the code, from how PAI reads a child’s message to how it decides what to do next.

The idea comes from Shariq’s work with more than 1,000 children, through over 100 hands-on workshops. He kept seeing the same thing: a child would look unable to learn something, and then make progress the moment the explanation, the pace or the activity changed. The hard part was never explaining. It was working out which change that child needed.

So PAI is designed to understand a learner before it teaches them. We’re starting narrow, with maths for 14 to 16 year olds preparing for IGCSE, and building toward younger, curious children. When a child says “I don’t understand fractions,” one child might think 1/4 is bigger than 1/3 because 4 is bigger than 3. Another might understand perfectly and just need more time. PAI asks first, makes a careful guess, watches what happens, and changes course if it got the child wrong.

Building that safely is my job. PAI isn’t a chatbot let loose on children. A rules engine decides what kind of help comes next: a prompt, a hint, a stronger hint, or a fresh explanation. The AI writes only within those limits, and a second model checks every reply before a child sees it, so it never simply hands over the answer. Maths answers are marked by exact arithmetic, not by the AI. It’s tested by more than 850 automated checks, and by red-team tests that actively try to make it fail.

We’re honest about where we are: PAI Learn is in development and hasn’t launched yet. We’re preparing our first pilot: putting the first version in front of a small group of learners, with their parents’ permission, to find out whether it really helps.

There’s something I find quietly remarkable about this one. The person who once stood in front of my class is now the person I’m building a company with.

“PAI isn’t a chatbot let loose on children.”

How it works

Installable web app (React, TypeScript, Vite) on AWS: Node.js on Lambda, PostgreSQL with pgvector, separate sign-in for students, parents and operators, all defined as code. AI runs through Amazon Bedrock. Claude writes within the rules engine’s limits, a second Claude model screens what a child types and checks every reply for answer leaks, and maths answers are marked by exact arithmetic, not by the AI.

  • React, TypeScript and Vite
  • Node.js on AWS Lambda
  • PostgreSQL with pgvector
  • Amazon Bedrock (Claude)
  • Playwright
    • Child types
    • Input screenClaude Haiku
    • Rules enginePicks the kind of help: a prompt, a hint, a stronger hint, or a fresh explanation
    • Claude SonnetWrites only within those limits
    • Answer-leak checkEvery reply is checked before a child sees it
    • Child

The proof

automated tests
850+
answer leaks in 20 attempts to force one
0

850+ automated tests, Playwright at phone, tablet and desktop sizes, and red-team suites against the real models. In our internal runs on the maths build, 20 attempts to make PAI reveal an answer produced 0 leaks, and 21 safety cases were handled with 0 missed and 0 false alarms.

What I learned

In a product for children, what the AI is not allowed to do matters as much as what it can do.

An idea that won’t leave you alone?