Practise and get graded
The student configures a mock test by subject, chapter, and topic, sits it with a timer and KaTeX-rendered questions, and AI scores the written answers with partial marks and feedback.
EdTech ecosystem · RAG AI tutoring · Live classes
Topkorbo
An AI-powered test-preparation and mentorship ecosystem for Bangladeshi students and teachers — mock exams with AI evaluation, a RAG-grounded reading companion, live mentor classes, contests, a community forum, and paid subscriptions in one platform.
Read the case study Visit the live siteTopkorbo puts the whole preparation loop in one place. A student takes a configurable mock exam and gets AI-graded feedback — including on written answers — then chats with a book-grounded AI tutor, joins a live class streamed by their mentor, competes in teacher-run contests, and discusses it all in a real-time community forum. Mentors watch each student's analytics from their own dashboard, and subscriptions with a free trial keep the platform sustainable.
What success needed to look like
- 01
Grade written and creative answers with AI so feedback arrives in minutes, not days.
- 02
Ground the AI book tutor in the student's actual textbooks through retrieval, not a model's memory.
- 03
Stream live mentor classes with screen sharing that survive refreshes and dropped connections.
- 04
Give mentors a dashboard view of each student's mock history, rankings, and weak subjects.
- 05
Run rated contests, a real-time community forum, and IELTS practice on the same platform.
- 06
Handle real payments, trials, and plan limits with SSLCommerz from day one.
In build
Actively in development on a staging deployment
How the solution works
From problem to outcome.
A simple view of how work moves through the solution, without exposing private implementation details.
Ask the book, not the void
In the reading library, the RAG pipeline retrieves the relevant passages of the uploaded textbook from vector search and the AI tutor answers from them — with one-click summaries, key points, and page-based quizzes.
Meet the mentor live
A connected mentor schedules a live class; the student joins with audio, video, and screen sharing, and can rejoin after a disconnect until the session ends.
Compete and discuss
Teacher-created contests rank participants, and the Socket.IO-powered forum carries the discussion — posts, threaded comments, reactions, and mentions land in real time.
Read the step-by-step workflow
- Practise and get graded: The student configures a mock test by subject, chapter, and topic, sits it with a timer and KaTeX-rendered questions, and AI scores the written answers with partial marks and feedback.
- Ask the book, not the void: In the reading library, the RAG pipeline retrieves the relevant passages of the uploaded textbook from vector search and the AI tutor answers from them — with one-click summaries, key points, and page-based quizzes.
- Meet the mentor live: A connected mentor schedules a live class; the student joins with audio, video, and screen sharing, and can rejoin after a disconnect until the session ends.
- Compete and discuss: Teacher-created contests rank participants, and the Socket.IO-powered forum carries the discussion — posts, threaded comments, reactions, and mentions land in real time.
The product in action
How BDBot works inside Discord.
Players never need to leave the community space they already use. The bot turns commands into clear match, profile, and leaderboard experiences directly in the channel.
Exam prep, framed as a sport
The homepage leads with competition: a rating card a student can imagine as their own, an arena to join, and mentors from top universities. Behind the navigation sit the AI tutor, live classes, contests, and the community that make it an ecosystem rather than a question bank.
The situation that needed to change.
Exam preparation in Bangladesh is fragmented: mock tests in one app, coaching classes in a physical room, mentors reachable only in person, and past papers in books. Feedback on written answers takes days, and a generic chatbot is no substitute — it answers from a model's memory instead of the student's actual textbook and question bank.
What I changed and built.
- Built AI into the grading loop, not just a chatbox: written and creative answers are scored automatically with partial marks, feedback, and worked mathematical explanations.
- Built a RAG pipeline for the reading library — teachers upload textbook PDFs, the server extracts and chunks the text, embeds it into MongoDB Atlas Vector Search, and the AI book tutor answers from retrieved passages of the actual book.
- Built live classes into the mentor relationship: mentors schedule sessions with camera, microphone, and screen sharing, and sessions survive refreshes and disconnections until the class ends.
- Made the whole platform real-time — forum posts, reactions, notifications, and class updates broadcast live over WebSockets.
- Wired in SSLCommerz payments so subscriptions, trials, and plan limits work end to end rather than as a mockup.
Key choices
Why this approach worked.
Retrieval before generation
The book tutor chunks uploaded PDFs, embeds them into MongoDB Atlas Vector Search, and answers from retrieved segments. A student revising for an exam needs the textbook's answer, not a fluent guess — and chat history stays isolated per book so context never bleeds.
AI grades the hard part
Multiple-choice is easy to score; written and creative answers are where feedback dies. AI evaluation returns partial marks, feedback, and worked mathematical explanations, so the slowest part of exam practice becomes the fastest.
Live classes that survive a bad connection
Sessions stay active through refreshes and disconnections until the mentor ends the class or time expires — because load-shedding and flaky networks are the reality the students live with.
Real-time by default
Forum posts, reactions, notifications, and class updates broadcast over Socket.IO rooms, with rate limiting on write operations so the community stays fast without drowning in spam.
Payments are a feature, not an afterthought
SSLCommerz checkout with success, failure, cancellation, and callback validation before any plan activates — plus a five-day Pro trial that locks premium tools cleanly when it lapses.
What I was responsible for.
I am building the platform end to end with my team: the mentor panel and live class system, the RAG-powered reading library, pricing and SSLCommerz payment integration, and the support system — alongside the mock test engine, AI evaluation, IELTS hub, contests, and community forum that complete the ecosystem.
- Build the mentor panel: student–mentor connections, dashboards, and per-student analytics.
- Build the live class system: scheduling, audio, video, screen sharing, rejoin support, history, and notifications.
- Build the RAG reading library: PDF extraction, chunking, vector indexing, and the book-grounded AI chat.
- Build pricing pages and SSLCommerz payment integration with trials and plan enforcement.
- Build the cross-panel support ticket system between users and admins.
- Deliver the wider ecosystem with the team: mock test engine, AI written evaluation, IELTS hub, contests, and the real-time community forum.
Outcome
What changed because of the work.
The platform is running on a staging deployment while features are hardened for launch: mock exams with AI evaluation, the RAG reading companion, live mentor classes, contests, the community forum, mentor dashboards, support ticketing, and paid plans with a five-day trial.
Mock exams with MCQ, CQ, and written answers graded by AI with partial marks
RAG book tutor over MongoDB Atlas Vector Search, isolated per book
Live mentor classes with screen sharing and rejoin support
Real-time forum with threaded comments, reactions, and mentions over Socket.IO
Teacher-run contests with rankings and cheating detection
SSLCommerz subscriptions with a five-day Pro trial
React 19, Node.js and Express, MongoDB, deployed on Vercel
What it took to deliver
EdTechAgentic RAGAI evaluationLive streamingReal-time systemsPayments
Clear about what can be shared.
Some client material and implementation details remain private. The numbers, responsibilities, and outcomes shown here are limited to what can be shared and verified.