AI / RAG
ShikkhaAI — Grounded AI Tutor for NCTB Class 9–10
Bangladesh's grounded AI tutor for SSC (NCTB Class 9–10). Every answer is taught, cited to the exact book and page, and never fabricated — ask in Bangla, English, or Banglish and get worked solutions, equation graphs, Mermaid diagrams, and board-style questions.
Public demo
4,551 — NCTB pages ingested & cited
The problem
SSC students in Bangladesh study from 18 official NCTB textbooks, but generic chatbots hallucinate answers, ignore the actual syllabus, and can't teach in natural Bangla — there's no trustworthy tutor grounded in the exact books the board examines.
The approach
A production RAG + agentic system grounded exclusively in the NCTB Class 9–10 corpus — 18 books, 4,551 pages ingested via PyMuPDF with a Bengali Tesseract OCR fallback, parent-child chunked, and embedded with OpenAI text-embedding-3-large (3072-dim) into Qdrant. Retrieval fuses dense vectors with BM25 keyword search via Reciprocal Rank Fusion, then an LLM listwise reranker orders context before generation. An orchestrator routes each turn to a specialist agent — Teacher, math solver, vision homework checker (gpt-4o), or exam generator — and every claim is cited [book, page].
The results
Grounded, citation-backed answers across all 18 NCTB subjects with inline equation graphs and Mermaid diagrams. Multilingual replies in Bangla / English / Banglish, voice questions via Whisper, and homework-photo grading. Firebase auth (Google + email) with per-student history over WebSocket-streamed responses. Deployed on Vercel + Render + Neon + Qdrant Cloud.
Stack: Next.js 14, FastAPI, Qdrant, Neon Postgres, OpenAI gpt-4o, Whisper, Firebase