Konze AI Tutor

Spoken responses are transcribed, then graded as text against exam rubrics rather than scored on audio features. Eighteen rubric prompts live in the database rather than in application code, so examiners can revise criteria without a deploy. Grading runs as a structured pipeline, producing component judgements rather than a single opaque verdict. Final scores are recomputed in Python from those components, so the number a learner sees never depends on model arithmetic.

That last decision is the one that makes the feature shippable. A language model is good at judging whether a response met a criterion and unreliable at totalling the result. Separating the judgement from the arithmetic means a grading error is a rubric problem you can inspect and fix, not a silent miscalculation in a learner's band score.

Four 7-day curricula covering IELTS General and Academic, and PTE Academic and Core, delivered as day-by-day lessons with slides, FAQs and exercises. Sixty-five generated lesson scripts underpin that delivery. Multi-section diagnostic testing places a learner before the curriculum starts.

Onboarding covers exam selection and personalisation, feeding a dashboard that tracks progress against the chosen track. AI practice spans all four skills — speaking, reading, writing and listening — each with automated evaluation.

Phase 1 · Foundations. Environment and repository setup, core data models, and account access — signup and login, JWT access and refresh authentication, OTP verification, Google OAuth — with onboarding screens and protected-route access control.

Phase 2 · Learning core. Onboarding flow with account creation, exam selection and personalisation. Dashboard and progress tracking. Curriculum delivery: generation, day-by-day lessons, slides, FAQs, exercises. Multi-section diagnostic testing and the initial PTE mock-test flow.

Phase 3 · AI practice & payments. AI practice across speaking, reading, writing and listening with automated evaluation, plus live conversational voice sessions. Subscription and payment handling. IELTS elevated to a first-class track with its own exam structure and mock-test engine.

Phase 4 · Hardening. Error monitoring and global error handling. Bug fixes across curriculum routing and AI-practice session timing, sourced correctly from the server. Separation of QA and production deploy pipelines, and deployment-process corrections.

Phase 5 · Production cadence. Steady release cadence with continued feature merges and fixes. Groundwork for activating the content retrieval layer already built but not yet live, and consolidating the platform's route and access surfaces.

Platform detail. Account access is built on JWT access and refresh tokens, OTP verification and Google OAuth, with protected-route access control throughout. Subscription and payment handling covers checkout, order tracking, payment history and add-ons. A vector-search retrieval layer over exam content is built into the platform, ready to activate.

What that changed is scope per unit of time. A team this size shipping a full authentication stack, four curricula, a diagnostic engine, two exam formats with separate mock-test engines, an automated grading pipeline, three conversational voice agents and a payments system, and then holding a steady production release cadence, is not the usual arithmetic. Claude Code was the reason it worked out, and it worked because the codebase was treated as the shared context rather than the prompts being treated as the work.

The grading pipeline is the clearest illustration. Getting from “grade this against a rubric” to a design where rubric prompts live in the database and final arithmetic is recomputed in Python is a series of engineering judgements, not a single generation. Working through those iterations quickly is what produced a design that could actually go to production.

Stack: Claude Code · React · TypeScript · Django · JWT · OAuth · OTP · vector search · conversational voice agents