Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Design for learning, not answers: the tutor guides students through problems step by step, grounded in their state board or national syllabus, in the language they study in. Build for low-end Android phones and patchy data, give parents a simple progress view, and measure success by improvement on practice tests rather than questions asked.
How to approach it
Pin down the user: a Class 10 student, often studying in Hindi, Marathi, Tamil or another medium, preparing for a high-stakes exam, sometimes with limited access to coaching. Clarify who pays (often parents) and the constraints: device, data, language, and the syllabus board (CBSE or a state board). Then name the tension: a model that hands out answers makes homework easier and learning worse. That tension has been measured. In a field experiment with about a thousand high-school maths students, those given unrestricted GPT-4 practice did better during practice but scored 17% lower on later exams without it; a version designed as a tutor, with teacher-written hints and safeguards against handing out answers, largely removed that harm (Bastani et al., PNAS, July 2025).
A strong answer
Users and jobs. The student wants to understand topics they find hard, practise exam-style questions, and know whether they are ready. The parent wants evidence their child is progressing and reassurance the app is safe. Teachers, where involved, want to know which topics a class is struggling with.
Core experience.
- Guided problem solving. A student photographs a maths problem. Instead of the answer, the tutor asks what they have tried, gives the next hint, and checks each step. A "show full solution" option exists after genuine attempts, because some students need a worked example to learn from.
- Syllabus-grounded explanations. Explanations draw on the student's textbook content and board's exam pattern, so terms, methods and examples match what the exam expects. Answers that use a method the board does not teach confuse students, even when correct.
- Exam practice. Past-paper style questions with marking schemes. The tutor grades written answers against the scheme and explains where marks were lost, which is what a board exam actually rewards.
- Readiness view. A topic map showing strengths and gaps, updated from practice results.
Language. Explanations in the student's medium of instruction, with technical terms in the form their textbook uses (often English terms within a Hindi explanation, for example). Voice input and output for students who find typing in their script slow.
The exam calendar changed in 2026. CBSE now runs two Class 10 board examinations, and "the better of the two performances is considered". In its first cycle, 5,13,955 students sat the second examination to improve, and 3,08,095 (59.95%) did (CBSE, 18 July 2026). For the product, that creates a second, sharply defined moment of need: a student who has just seen their first result and has weeks to fix specific weak topics. A readiness view built from the first paper's marks is the feature that period wants.
Constraints that shape the build.
- Devices and data. Many students use shared, low-end Android phones with limited data, and connectivity at school is far from universal: the share of schools with internet rose from 63.5% to 67.4% in 2025-26 (UDISE+ via PIB, July 2026). Keep the app small, text-first, with images compressed, and let students download a chapter's practice set for offline use.
- Cost. Per-student inference cost must fit a price parents can pay, so routine hints run on a small, cheap model, with a stronger model for grading long answers or explaining difficult concepts.
- Accuracy. A wrong explanation in maths or science teaches a wrong method. Maths answers are checked by computation where possible; content is evaluated against textbook solutions by subject experts before each subject launches.
Safety. Students are minors. No open-ended social chat, content filters tuned for school-age users, and data practices that respect parental consent under India's DPDP rules, whose provisions on children's data apply from about 13 May 2027 (as of November 2026; confirm with counsel). Building verifiable parental consent before that date is cheaper than retrofitting it.
Metrics. Improvement on practice test scores over a month (the outcome), share of problems solved with hints rather than full solutions, weekly active study days, parent-reported confidence, and, as guardrails, error reports on explanations and share of sessions that go straight to "show solution".
What interviewers probe next
"Students just want answers. Won't they leave for an app that gives them?" Some will. Keep the worked solution available after an attempt, and show progress on practice tests, which is what students and parents care about at exam time.
"How do you launch across many state boards?" Start with CBSE maths and science in Hindi and English, where demand and content are largest, then add boards and languages one at a time, each with expert evaluation.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Learning design | Answers questions | Step-by-step help | Hints first, worked solutions after attempts, grading against the board's marking scheme |
| Context | Generic global app | Mentions languages | Medium of instruction, syllabus grounding, low-end devices, offline use, price-driven model choice |
| Measurement | Questions asked | Engagement | Practice-score improvement with hint-versus-solution ratio and accuracy guardrails |
Common mistakes
Building an answer machine that undermines learning.
Ignoring syllabus and marking scheme, which decide exam results.
Designing for a flagship phone and unlimited data.
See the AI fit test and cost of a wrong answer.