Landing Page
Student Dashboard
Student Learning Path
New Lesson Creation
Student My Lessons
Practice Lab Simulation
Practice Lab Coding Challenges
Student Progress Page
Student Profile
Admin Dashboard
Admin Mentor Inbox
Admin Learning Insights
Admin View Users
Admin System Health
Admin Profile

Inspiration

Students often study using PDFs, slides, and other course materials scattered across different sources. General-purpose AI chatbots may provide answers that are not grounded in the actual course content. We built Shikshak AI to provide a more reliable and personalized learning experience based on students' own study materials and their understanding of each concept.

What it does

Shikshak AI transforms uploaded course materials into a structured learning experience. Its key features include:

Course Material Processing: Upload PDF, DOCX, PPTX, TXT, or Markdown files to generate a syllabus of topics and concepts.
Source-Grounded Tutoring: Get answers with page-level citations and refusal of unsupported queries.
AI Video Lessons: Learn through an AI avatar with whiteboard slides and English or Hindi text-to-speech.
Adaptive Assessments: Take placement checks and quizzes with answer evaluation, feedback, and misconception reports.
Personalized Learning: Track per-concept mastery through a skill map and revisit concepts using spaced review.

How we built it

We built the backend using FastAPI and the frontend using Next.js 16, React 19, and TypeScript. Google Gemini provides core language-model functionality, while BGE-M3 supports hybrid dense and sparse retrieval with reranking over ChromaDB. The system architecture relies on a multi-agent pedagogical state machine that custom-coordinates the planner, teacher, and evaluator. SQLite with write-ahead logging (WAL) supports mastery tracking, learning paths, and reviews. The project is fully Dockerized and includes a fallback mode that operates without API keys. Our team built the core architecture and multi-agent system, using ChatGPT and Gemini purely for ideation, prototyping, and coding support.

Challenges we ran into

One of our main challenges was keeping answers grounded in uploaded course material while maintaining a natural, conversational tutoring experience. We also worked on synchronizing avatar videos, text-to-speech, and whiteboard slides, along with handling Hindi text rendering on generated whiteboards.

Accomplishments that we're proud of

We brought together source-grounded tutoring, AI avatar lessons, adaptive assessments, and per-concept learning progress in a single platform. We are particularly proud of combining cited answers and unsupported-query refusal with a learning workflow that helps students assess their understanding and revisit concepts.

What we learned

We learned how to integrate retrieval-augmented generation with source citations, coordinate a multi-agent tutoring workflow, and connect assessment feedback with learner progress. We also gained experience integrating avatar-based teaching, text-to-speech, and multilingual support into a unified learning application.

What's next for Shikshak AI

Add lecture-video ingestion with timestamp-based citations.
Benchmark retrieval and generation using RAGAS on a public course.
Run simulated-student experiments to measure mastery gains and question repetition.
Develop a visual map of topics and prerequisites.