Bangladesh's First Integrated AI Learning Ecosystem
A 10-month journey building TenTen inside the 10 Minute School app: from early R&D in Google Colab to production RAG, Bangla LLM behaviour, intent routing, and product analytics after a limited release in Nov '25.
When learning paused, students had nowhere to turn
10 Minute School reaches millions of learners across Bangladesh, many of them without access to private tutoring. Students on the platform could learn, but they had no way to recover when they got confused.
Four issues kept showing up:
Stuck watching a recorded class. No teacher available to ask for clarification.
Stuck after reading an exam explanation that still didn’t make sense. Stuck with no one to ask.
Stuck on a theory question with nowhere useful to look for a real explanation.
Stuck during live class, with doubts piling up and platform help buried in FAQ pages nobody reads.
That meant the product was good at delivering content but had no answer for the moment after confusion. And in education, that moment matters more than almost anything else. It is where students pause, skip, guess, or quietly give up.
Before TenTen, support meant FAQ pages and external chatbots with zero curriculum context. Learning broke at the exact moments students needed help most, and nothing inside 10MS could stitch it back together.
The solution
Not a regular chatbot
Solving academic challenges inside the learning journey
TenTen lives exactly where learning breaks: inside general chat, recorded lessons, exam review, and live class. It is built into 10 Minute School, not bolted on as a separate chat window, so help arrives with the student's class, grade, and curriculum context already in place.
Solve problems on demand
Study partner, not teacher. TenTen does not lecture. It helps students work through confusion the way a smart classmate would: step by step, in Bangla when they need it.
Ground answers in what they’re learning
Academic doubts pull from textbooks and trusted 10MS content. Platform questions route elsewhere. Students get the right kind of help instead of one generic reply for every intent.
Grow alongside the learner
As students use TenTen across surfaces, it stays aware of their progress: what they watched, what they got wrong, what to revise next. Guidance deepens instead of resetting every session.
With TenTen, the product finally supports the moment after confusion, not just content consumption. A study partner embedded in the learning journey, not another chatbot students have to leave the app to find.
What I built
Four systems shipped across a ten-month build: retrieval, routing, orchestration, and analytics.
From pilot to product
Where the engineering work fit in TenTen's story
The work covered the data and AI layer: retrieval, routing, and orchestration that turned early Colab, n8n, and Dify experiments into production systems the product could actually trust.
Eight phases from research through limited release. The cards below are the four pillars that shipped across that arc.
R&D
Prototype to production orchestration
~10 months of retrieval and prompt experiments in Colab, n8n, and Dify before production wiring.
Migrated orchestration to LangGraph for agent flows, state, and production control.
Colabn8nDifyLangGraph
RAG
Curriculum-grounded retrieval
Evidence from textbooks and trusted 10MS content, with eval loops on retrieval quality.
Semantic chunking, hybrid search (semantic + keyword), and reranking before generation.
Semantic chunkingHybrid searchReranking
Routing
Mixed-intent handling
Separate paths for academic doubts vs platform and logistics questions.
Academic and general agents, a 10MS MCP server, and subject-wise user memory retrieved when needed.
Multi-agentMCPUser memory
Analytics
Post-launch measurement
Data-led iteration after limited release in Nov ’25.
Usage, retention, session depth, and answer helpfulness fed back into product and model improvements.
UsageRetentionHelpfulness
Engineering challenges
Why those systems were necessary: the constraints and failure modes that shaped every decision, not just an implementation checklist.
অ
Bangla in the wild
Students rarely write textbook Bangla. Code-mixed shorthand, typos, and spoken phrasing break assumptions baked into base models and clean training data.
A fluent-sounding wrong answer erodes trust faster than an honest “I’m not sure.”
?
Sparse questions, dense books
A six-word doubt has to land in the right NCTB paragraph inside a long chapter. Vector similarity alone often returns plausible-but-wrong passages.
Retrieval mistakes look like knowledge mistakes to a student who can’t verify the source.
◎
One chat, many jobs
The same input handles physics steps, class timing, pricing, and platform help. Routing errors send academic answers to logistics questions, or flip it the other way around.
Wrong intent is worse than no answer when students are already stuck.
How TenTen shows up for students
Four learning surfaces, one integrated study partner across the moments that matter most. Not four separate chatbots.
General chat
24/7 doubt solving
Ask any academic question, anytime. Step-by-step help across every subject.
Live class
Doubts during and after class
Learning support inside live sessions. Ask a question without breaking the flow.
Recorded class
Help without leaving the video
Ask about a specific moment in a recorded lesson. Context travels with the question.
Exam solution
Understand why an answer was wrong
Instant explanations after exams. Follow up on any MCQ you got wrong and actually understand why.
Impact in numbers
Numbers that came back stronger than expected after just months of rollout.
0
Adoption rate
After multiple rollout phases, more than half of paid users have used TenTen.
0
Messages answered
Doubts and follow-ups handled across general chat, live class, recorded lessons, and exam review.
0
Daily active learners
On average, 500+ students actively solve doubts with TenTen each day.
Sales impact
After launch, 19.2% of students credited TenTen as a key reason for purchasing, and 50% said it will influence their next purchase too.
Student usage
Math leads, but practice, theory, and exam prep are close behind
How Students Use TenTen
Math Problem Solving
55.6%
MCQ & CQ Practice
50.6%
Theory & Explanations
49.6%
Step-by-Step Solutions
40.5%
Answer Checking
39.4%
Study Routine Planning
26.0%
Exam Prep & Quiz
23.6%
Class Routine Checking
19.2%
0%10%20%30%40%50%60%70%
% of Students
Reach
4,200+ locations across all eight divisions
AI reaching every corner of Bangladesh
0Students from rural areas
What this means
More than a tool. Part of how a generation learns.
Solving the basics is only the starting point. As the largest edTech platform in Bangladesh, there is a responsibility to bring bold innovations that actually reshape how students learn. That means reaching every division, from rural classrooms to city prep centers, with AI that grows alongside how students study.
Need help building an AI system that’s actually helpful?
From RAG and intent routing to launch analytics, happy to help teams ship AI that students actually trust inside real products, not just one-off demos.
R&D