AI Tutor & Learning App

What AI tutoring does well and where it falls short

Being specific about this early stops a client budgeting for features that will disappoint.

Teaching, not answering

The default behaviour of a language model is to answer the question. For a homework app, that is the worst possible behaviour — it produces a student who finishes their homework and learns nothing, and a parent who cancels the subscription when the exam results arrive.

• Ask before telling. "What have you tried?" then a hint, then a bigger hint, then the method, and only then the answer.
• Work one step at a time in maths. The student attempts each step; the tutor confirms or corrects it.
• Diagnose the misconception. A student who keeps dropping the negative sign has a specific gap, and naming it is more useful than correcting the arithmetic ten more times.
• Revisit deliberately. Anything answered wrong reappears three days later, then a week later. Spaced repetition is unglamorous and works.
• Report to parents in plain terms. Weak on quadratic equations, strong on trigonometry, forty minutes this week. That is the report a parent renews on.

Handling the photo of a handwritten answer

Students photograph their answer sheet and expect it marked. This works better than people assume and worse than a demo suggests.

Neat handwriting on ruled paper in good light reads well enough to mark reliably. Cramped writing, pencil, poor light, or a page photographed at an angle produces a transcript with enough errors that marking becomes unfair — and a student marked down for the app's misreading loses trust immediately.

The design answer is confidence-aware: mark what reads clearly, show the student the parts the app could not read and ask them to type those, and never award a low mark on an uncertain transcript. For diagram-heavy subjects, the honest position is that the app checks the written parts and leaves the diagram to a teacher.

Board coverage as a business decision

Building for CBSE first is usually right. It is the largest single syllabus, the material is well documented, past papers are abundant, and the same content covers a national audience.

State boards are a different project each time — different textbooks, different languages, different exam formats. Maharashtra State Board in Marathi is not a translation of CBSE in English; it is a separate corpus.

ICSE sits in between, with a smaller but higher-paying audience and a heavier emphasis on written expression, which raises the marking difficulty. We would launch on one board, prove that students improve, and expand from evidence rather than building three at once and doing all of them thinly.

Who pays, and what that changes

Students use the app; parents pay for it. Those are different people with different criteria, and an app designed only for the student does not convert.

The parent needs visibility — what has been studied, what is improving, what is not — delivered without being asked. A weekly summary message is the single most effective retention feature in this category.

The alternative model is selling to schools and coaching institutes, where one decision covers hundreds of students and churn is annual rather than monthly. That build needs a teacher dashboard, class assignment and bulk reporting, which is meaningfully more work but a far better business if you can reach the buyers.

What we deliver

The signed Android app, the syllabus corpus and retrieval layer on your infrastructure, the tutoring behaviour and marking rubrics as editable configuration, parent and student accounts with a compliant consent flow, progress analytics, and full source.

We also deliver the content pipeline — how new chapters, papers and marking schemes get added without an engineer. Syllabi change, and an app that needs a development cycle to absorb a revised chapter will fall behind within two years.