AiAssessment IntelligenceAI recommends. Teachers decide.

Technical brief

How this is built, and how it grows up

The prototype is deliberately boring infrastructure around one sharp AI capability. This page documents the architecture as-is and the credible path from vibe-coded prototype to a production ML system — the answer to the Board's "what would it take for real?"

Architecture today (five thin layers)

InterfaceNext.js / React + Tailwind, deployed on Vercel
OrchestrationRubric-driven prompt templating, single batched call, defensive JSON parsing
AI coreLLM classification of open-text responses into misconception clusters (temperature 0, JSON mode)
Human-in-the-loopApprove / Modify / Reject gate before any action; every decision logged
DataSynthetic demo dataset + cached reference analysis; localStorage decision log

Key design decision: the model is never trusted with arithmetic. src/lib/aggregate.ts recomputes every number deterministically, which is why each figure on screen is defensible under CFO questioning.

The ML roadmap: prompting first, training when it earns its keep

The honest sequencing: a prompted frontier LLM is the right Phase-1 tool because it works with zero training data. Dedicated trained models enter exactly when labelled data and volume make them better and cheaper — not before.

NOW · prototype

Prompted LLM classification

A frontier LLM (Gemini; swappable via one constant in src/lib/constants.ts) classifies each response against the rubric in one batched, temperature-0 call. The server recomputes every aggregate — draft marks, cluster membership, gap-map percentages — so no number on screen comes from model arithmetic. Synthetic data only; zero PII.

PHASE 2 · pilot

Grounding with real signals

Add few-shot exemplars per subject from teacher-verified classifications, and start capturing the Approve/Modify/Reject signal plus per-response corrections as labelled training data. Move storage from localStorage to a proper store (Supabase/Postgres or Azure Cosmos DB) with consented, de-identified responses. Add batch processing for full mock-test uploads.

PHASE 3 · production ML

Trained models where they beat prompting

Once thousands of teacher-verified labels exist, train a dedicated classifier for the frequent misconception categories: cheaper, faster, and more consistent than prompting for high-volume subjects. Keep the LLM for the long tail (novel misconceptions, explanations, intervention drafting). Embedding-based clustering surfaces new misconception types before anyone names them.

PHASE 4 · scale

Per-subject, multilingual, measured

Per-subject model variants (physics numericals need different signals than economics prose), regional-language support aligned to Project Bharat, and outcome tracking that ties approved interventions to score movement — closing the Measure step of the loop and feeding the ROI model presented to the Board.

Production platform choices (Azure track, with Google equivalents)

Managed LLM

Azure OpenAI Service (or Google Vertex AI). Enterprise SLAs, regional data residency (India regions), content filtering, and private networking — the compliance posture an education company listed in India needs. Our provider call is isolated in one file, so this swap is a config change, not a rebuild.

Custom classifier

Azure AI Language custom text classification, or Vertex AI AutoML text — train the Phase-3 misconception classifier on teacher-verified labels without building training infra. Export metrics per class to watch for drift.

MLOps loop

Azure Machine Learning (or Vertex Pipelines) to schedule retraining on fresh teacher feedback, evaluate against a held-out golden set, and gate deployment on accuracy + fairness checks across centres and languages.

Data platform

Consent-tagged, de-identified response store (Azure Cosmos DB / Postgres) with rubric standardisation as a first-class schema. The teacher decision log becomes the most valuable table in the company: it is labelled training data nobody else has.

Security & governance

Keys in managed identity / Key Vault (never in code — the prototype already enforces env-var keys), audit logs on every AI call, role-based access, and the human-approval gate preserved as an architectural layer at every phase.

What production needs that this prototype does not have

  • • Consented, de-identified real student response data (prototype: synthetic only).
  • • Standardised rubrics across centres and faculties.
  • • Outcome history to validate that approved interventions actually move scores.
  • • A fairness audit across student groups before any draft-mark feature ships wider.

Stated openly by design: the evaluation brief rewards honesty about limitations over overclaiming.

← How to use·Open the app →