The problem
Exam preparation platforms run out of questions. A fixed bank gets memorised, and the students who work hardest exhaust it fastest - which is the opposite of what should happen. Writing new papers is expensive and slow, so most platforms recycle, and a recycled paper stops measuring anything after the second attempt.
The second problem is the answer key. A student who gets a question wrong needs to know why, and a one-line key does not tell them. Written explanations for every question in a large bank is a content budget most platforms do not have.
What the model does
Questions are generated against a topic and a difficulty band rather than retrieved from a fixed bank, so the paper is new each time and difficulty tracks the student's own performance history. Each question comes with a worked explanation generated alongside it, which is the part that makes a wrong answer useful.
The engineering problem here is not prompting, it is validation. A generated question with four options and no correct answer, or two correct answers, is worse than no question at all. Every generation is parsed into a strict structure and checked before it is ever shown, and anything that fails is regenerated rather than displayed.
- Generation targets a topic and difficulty band, not a fixed question bank
- Worked explanations are produced with the question, not retrieved separately
- Strict schema validation rejects malformed questions before display
- Difficulty adapts from per-topic performance, tracked per user
Beyond question generation
The same generation and validation layer carries three further features, which is the argument for building the plumbing properly the first time. Resume analysis reads a student's CV and scores it against a target role, returning specific weaknesses rather than a grade. Interview practice generates role-appropriate questions and follow-ups that respond to the answer given. Performance heatmaps turn the per-topic history into something a student can act on.
None of these needed a new AI integration. They needed a different prompt and a different output schema against an interface that already handled retries, validation and cost tracking.
Why Groq
Latency is the product here. A student generating a practice paper will not wait thirty seconds, and a platform that feels slow does not get opened the following evening. Groq's inference speed on Llama-3 was the deciding factor over more capable but slower alternatives - for structured question generation against a well-specified schema, the ceiling on model capability was not what limited quality.
Cost per generation mattered for the same reason. A free-tier student generating unlimited papers is only viable if each paper costs a fraction of a rupee.