EduJha
An AI-assisted platform for India's competitive government exams. The interesting engineering is not the model call — it is everything built around it so that generated content can be trusted, papers can be specified rather than hand-built, and no single provider holds the product hostage.
- Competitive exam preparation runs on volume of practice and speed of feedback. Static question banks give neither — they don't adapt to where a candidate is weak, and human evaluation can't answer while the attempt is still fresh.
- A platform, not a quiz app. Timed mock tests scored in real time, an AI tutor grounded in the candidate's own performance history, and analytics that turn each attempt into the next recommendation. Model calls run across several providers behind one interface, so no single vendor sets the cost or latency ceiling.
- A production web platform across multiple exam tracks — timed mock tests, immediate scoring, an AI tutor, per-candidate analytics, and subscription billing — with a native Android and iOS app built on the same backend.
Exams as a validated specification
Papers are declared as blueprints rather than hand-assembled — question types, scoring strategy, difficulty distribution, selection rules, randomisation, and navigation constraints, expressed as a recursive schema and validated before anything reaches a candidate. Adding an exam track becomes a specification, not a code change.
A generation pipeline with a validation gate
AI-generated questions pass through an explicit pipeline that separates generation from validation, so output is checked against the blueprint's constraints before it can be published. The failure mode this prevents — plausible-looking questions that break the paper's own rules — is the one that matters at scale.
Provider-portable model layer
Anthropic, Google, and Groq models sit behind one interface. Routing is a configuration decision, which keeps cost and latency negotiable and removes the single-vendor dependency most AI products accumulate by default.
Versioned migration in production
A bridge service lets a newer content model and its predecessor run side by side, so the platform migrated forward without a freeze — the same strangler approach we apply to enterprise modernization.
Abuse and access control
Rate limiting and access control sit in front of the expensive paths, because inference cost is an attack surface as much as a line item.
Built to be found
Canonical URL handling, programmatic internal linking, a segmented sitemap index, and an llms.txt for AI crawlers — discovery treated as engineering rather than an afterthought.
One backend, web and mobile
A React Native app built with Expo shares the platform's backend: Firebase sign-in, Razorpay payments, push notifications, secure token storage, and an offline SQLite cache so candidates can keep practising on a patchy connection.
- Next.js (App Router)
- TypeScript
- Prisma / PostgreSQL
- NextAuth v5
- Anthropic, Google & Groq models
- Vercel AI SDK
- React Native (Expo)
Want this kind of engineering on your system?
The team that built this takes on client work. Tell us what you're trying to change.