Verity AI

One truth: your teacher's material, in every student's language.

A closed-corpus AI learning platform for students learning in English as a Second Language (ESL) at IB & Cambridge international schools. Every answer comes only from a teacher's own approved material — never the open internet — cited every time, and translated into a student's own language.

Live Demo →  ·  Pitch Deck →  ·  Demo Script →

Built end-to-end for the Claude Code Hackathon

---

Why Verity AI

Generic AI tutors answer from the whole internet — unpredictable, uncited, and easy for a student to use to just cheat. Verity AI does the opposite: it is architecturally constrained to answer only from material a teacher has approved, cites the exact source every time, and is designed to guide a student toward an answer rather than hand one over.

🔒 Closed-corpus, cited	Every answer traces to an exact slide or worksheet — no open-internet knowledge, no hallucinated sources.
🌏 Multilingual by design	Explains and translates into a student's own language — Chinese (Mandarin, Simplified) live today, Korean/Malay/Tamil next.
🎯 Guides, never cheats	The AI hints and asks Socratic questions — it will not complete a student's assignment for them.
✅ Trusted grading	Numbers and multiple-choice are graded by deterministic, unit-tested code — never by an LLM guessing.
📈 Adaptive & transparent	Difficulty adapts to the learner; every AI conversation is fully visible to the teacher.

Contents

• How it works
• What's live today
• Architecture
• Tech stack
• Quick start
• Testing
• Roadmap
• Team

How it works

flowchart LR
    A["Teacher's approved material<br/>(slides + worksheet answer keys)"] --> B["Closed corpus<br/>(vectorless, long-context RAG)"]
    Q["Student question"] --> C{{"Claude"}}
    B --> C
    C -->|"Explain · Example · Ask Me · Check · Translate"| E["Cited answer<br/>Based on: &lt;exact source&gt;"]
    E --> G["Deterministic grader<br/>(numbers & MCQ only — never the LLM)"]

A student never has to write a prompt — they tap one of five intents (Explain, Give Example, Ask Me Questions, Check My Answer, Translate), and the system prompt architecturally forbids the model from using anything outside the approved corpus or completing an assignment on the student's behalf.

What's live today

Two complete Physics topics, built from the school's real Grade 7 teaching material (not synthetic content):

• Moments of a Force — interactive drag-to-balance seesaw simulation, moment/torque worked examples.
• Distance–Time Graphs — a scrubbable journey graph syncing a moving figure to a live distance-time plot.

Each topic ships with the full 5-mode AI Learning Assistant, a deterministic auto-grading Practice Zone (unit-tested), and an ESL Reading Assistant (inline glossary + text-to-speech, bilingual English/Chinese).

Four role-based views, matching how a real school is structured:

Role	View
🎒 Student	The AI Learning Assistant, interactive visuals, and adaptive practice
👩‍🏫 Teacher	Full AI-chat transparency, reliance flags, class progress
📊 HOD	Department-level rollup across sections and teachers
🏫 Principal	School-wide completion, engagement, and ESL-improvement dashboards

Honest note on data: the AI learning assistant, grading, and translation are fully live and functional. The Teacher/HOD/Principal dashboards currently run on realistic, clearly-labeled illustrative data — real authentication and persistence (Supabase) is the next build phase, not yet wired up in the hackathon window.

Architecture

• Retrieval is hybrid by design. A small per-topic corpus is injected directly into Claude's context as vectorless, long-context RAG — no embedding drift, maximum accuracy at classroom scale. The architecture is designed to graduate to pgvector for whole-school libraries, and a curriculum knowledge graph later, for adaptive learning paths.
• Accuracy is deterministic wherever it can be. Numbers and MCQs are graded by rule-based code in src/lib/grade.ts — unit-tested, never left to an LLM to "guess" a grade. The LLM is reserved for what LLMs are actually good at: explanation, Socratic questioning, translation, and rubric-based feedback.
• Academic integrity is enforced in the system prompt itself (src/lib/tutor.ts) — the AI is architecturally forbidden from completing an assignment, and always cites its source.
• Translation is corpus-grounded, not generic MT. Every Chinese translation ties back to the exact approved-material chunk it came from (src/data/translations-zh.ts), reviewed for terminology consistency.

Tech stack

Next.js 16 (App Router) · TypeScript · Tailwind CSS v4 · Framer Motion · Anthropic Claude API · Vitest · Supabase (auth/roles/storage/pgvector — planned) · Deployed on Vercel

Quick start

npm install
cp .env.local.example .env.local   # add ANTHROPIC_API_KEY for the live AI
npm run dev                        # http://localhost:3000

Without an API key the app runs in demo mode with curated, still-cited fallback responses, so the whole flow is clickable end-to-end.

Testing

npx vitest run    # deterministic grader — value, unit, direction, tolerance

Roadmap

Teacher material upload with auto-question generation · real Supabase authentication & persistence · Korean, Malay, and Tamil language packs · additional subjects beyond Physics · whole-school vector + graph RAG · native iPad app.

Team

Manoj Bhardwaj — Founder & Builder — 20+ years in enterprise banking technology at global financial institutions, building systems that must be correct, auditable, and trusted at scale. Currently founder of Dhari AI, building production-grade reasoning agents for regulated institutions with banking-grade rigor — the same discipline applied here to ESL education.

License

MIT — see the license file for details.