How to Build an AI Tutoring or Learning Platform: Lessons From PadhAI

To build a learning product, get assessment, AI cost, accessibility, student privacy and exam-time reliability right before engagement features. RAITHub built PadhAI, an AI tutoring platform with adaptive assessment, a Socratic tutor, a cost-aware model router and the same experience across a PWA, WhatsApp and Telegram, and this guide draws on it.

It is written for founders and product leads planning a tutoring app, a course platform, an assessment engine or a learning management system (LMS). It covers what to get right and where the hard problems are. If you are looking for a team to build yours, the service page is EdTech software development.

What edtech software has RAITHub built?

PadhAI, an AI-powered personalised tutoring platform designed for 7 emerging markets. RAITHub built it, and its case study is available as a PDF.

Part of PadhAI	What it does	Why it matters for a learning product
Polyglot services	11 services: 9 Node/TypeScript and 2 Python/FastAPI	Product logic and AI or data work each run in the language suited to them
70/20/10 model router	Sends each query to a lightweight, mid-tier or premium model by complexity	Keeps AI cost per student viable
Adaptive assessment	Knowledge tracing estimates what each learner has mastered	Questions and practice follow the learner, not a fixed sequence
Socratic tutor	Guides with questions, uses math verification and retrieval-augmented generation (RAG)	Checks answers instead of trusting the model, and grounds replies in source material
Channel parity	The same learning experience across a PWA, WhatsApp and Telegram	Learners use the apps they already have
Payments	9 payment gateways under one abstraction	A new market's gateway is an adapter, not a rewrite

This page does not quote PadhAI user numbers, launch dates or learning outcomes. The architecture is described on the work page, and the full write-up is in the PadhAI case study (PDF). RAITHub's test discipline is shown in its commerce platforms: 750+ automated tests on TheSkinProof (the founder's own marketplace, built and run by RAITHub) and 530+ on Sundor Skin.

What does edtech software development need to get right?

Five things: assessment that measures learning correctly, content and progress that teachers and learners trust, accessibility, student-data privacy, and reliability at exam time. Engagement features only work on top of those.

The product shape changes the emphasis. An LMS is mostly about courses, enrolment, grading and integration with institutional systems. An e-learning app for individual learners is mostly about onboarding, habit and payments. An AI tutoring app adds model cost, answer correctness and safety. The sections below take each concern in turn.

How should assessment and progress tracking work?

Record every attempt as an event, tag every question with the skills it tests, and derive progress from that history. Never store progress only as a number that code overwrites.

Knowledge tracing is a family of methods that estimate, from a learner's sequence of answers, how likely they are to have mastered each skill. It lets a product pick the next question sensibly: not so easy that it wastes time, not so hard that the learner quits. It also depends on a well-tagged item bank. A question tagged with the wrong skill quietly corrupts every estimate built on it.

Progress also has to be explainable. A teacher or parent who asks "why does it say she has mastered fractions?" deserves an answer drawn from real attempts. Grading logic deserves the strictest tests in the codebase, because a scoring bug is a fairness problem, not a cosmetic one. Streaks, badges and leaderboards sit on top of this, and only work when the progress data underneath is right.

How do you keep AI tutoring app costs under control?

Route each query to the least expensive model that can answer it well, and put hard limits around the rest. Sending every question to the most capable model is the fastest way to make a tutoring product uneconomic.

PadhAI's 70/20/10 model router classifies query complexity and sends each query to a lightweight, mid-tier or premium model. Other controls worth designing in from the start:

• Per-learner budgets and rate limits, so one heavy user or a script cannot run up the bill.
• Caching of answers to common, non-personal questions.
• Shorter context: send the model the relevant material, retrieved by RAG, rather than everything.
• Cost per learner as a tracked metric, visible to the team every week.

How do you make an AI tutor trustworthy?

Do not let the model be the only judge of correctness. Verify what can be verified, ground answers in approved material, and design the tutor to teach rather than hand over answers.

PadhAI's tutor is Socratic: it guides the learner with questions instead of giving the solution. It uses math verification, so a mathematical result is checked rather than accepted on the model's word, and RAG, so explanations draw on source material. For younger learners, add content filtering, clear escalation to a human, and logs a safeguarding lead can review.

How do LTI and SSO integrations work?

LTI lets an LMS launch your tool and pass identity, roles and grades securely; SSO lets users sign in with their institution's identity provider. Schools and universities often expect both.

According to 1EdTech, which maintains the standard, LTI 1.3 uses a security framework based on OAuth 2.0 and JSON Web Tokens. LTI Advantage adds three services: Assignment and Grade Services, Names and Role Provisioning Services, and Deep Linking. SSO usually means SAML or OpenID Connect against the institution's identity provider. LTI and institutional SSO are not part of PadhAI's published description, so on your project they would be new work, estimated and tested as such.

What does accessibility require in an edtech product?

Treat WCAG as a working practice from the first screen, not an audit at the end. Education buyers often ask about accessibility in procurement, and learners with disabilities cannot use what they cannot operate.

The W3C describes WCAG 2.2 as the current version, with three conformance levels: A, AA and AAA. In practice that means semantic HTML, full keyboard operation, visible focus, sufficient contrast, captions and transcripts for media, and forms that work with screen readers. Automated checks catch part of this; manual testing with assistive technology catches the rest. RAITHub works to WCAG as a practice. It does not certify conformance, and if you need a formal conformance report, an independent accessibility auditor should produce it.

How do you protect student data?

Collect less, keep it for less time, control who can see it, and find out early which privacy laws apply. Children's data carries extra rules in many jurisdictions.

Rule	What it covers (in general)	What it usually means for engineering
COPPA (United States)	Operators of websites or online services directed to children under 13, or with actual knowledge that they collect personal information from a child under 13	Age screening, verifiable parental consent flows, and limits on what is collected
GDPR Article 8 (European Union)	Where consent is the basis for an online service offered directly to a child: parental consent below 16, which member states may lower to no less than 13	Age thresholds configurable by country, and consent records you can produce
Other national and state laws	Varies, including student-record and school-procurement rules	Confirm with counsel before choosing data stores, vendors and AI providers

Two engineering points apply almost everywhere. Check the data terms of every AI provider before student text reaches it. And build deletion and export into the data model from the start, because adding them later is slow.

Not legal advice. These are general summaries. Which laws apply to your product, and what they require, is a question for a qualified privacy lawyer in each market you serve.

How do you prepare for exam-time traffic peaks?

Load-test for the peak you expect, move slow work onto queues, and decide in advance what degrades gracefully. Enrolment days and exam weeks concentrate traffic in a way ordinary days do not.

• Test the real peak. Simulate the exam-hour pattern: many learners starting at once, submitting at once.
• Queue heavy work. Grading, report generation and notifications run as background jobs, not inside the request.
• Protect submissions first. Saving an answer must work even if the AI tutor or analytics is slowed or paused.
• Watch it live. Alerts on error rate and latency, with a runbook for the person on call.

RAITHub's QA & Reliability retainer covers this kind of work on an ongoing basis; see the services page.

In what order should you build a learning product?

Data model first, engagement last. The attempt history is what every later feature reads from, so it has to be right before anything is built on it.

1. Learners, content, attempts and progress. The data model, with every attempt stored as an event.
2. Assessment and grading. The strictest tests in the codebase live here.
3. The tutor or course experience, with cost limits and verification from the start.
4. Payments and access, if learners or schools pay.
5. Engagement features, once the learning loop works.

A focused first version typically takes 4–6 weeks; the cost section above covers budget drivers, and the MVP cost estimator gives a quick range. RAITHub's phases are in how RAITHub delivers software, the testing method in how RAITHub tests, and how quotes work on the pricing page.

What it costs to build a learning platform in 2026

Cleveroad's June 2026 guide puts a basic learning app MVP at $40,000–$70,000 or more, a mid-complexity platform at $70,000–$150,000 and a full-featured LMS at $150,000–$250,000 or more. Those are market figures, not RAITHub quotes. RAITHub publishes no rates; it gives a written fixed-scope estimate after a free 15-minute technical audit.

Learning products have cost lines that most web apps do not, and one of them keeps running after launch:

• AI model spend. An AI tutor has a running cost per learner, per question. The router, caching and per-learner limits described above are build work that decides whether that running cost is viable at your price.
• Assessment and the item bank. Knowledge tracing only works on a well-tagged item bank, and grading logic needs the strictest tests in the codebase.
• Channels. A PWA, WhatsApp and Telegram are three front ends with different message formats and limits. WhatsApp's business platform also has its own approval rules and fees.
• Institutional integrations. LTI 1.3 launches, grade passback and institutional SSO, each tested against the specific LMS and identity provider your buyers use.
• Children's data. Age screening, parental consent records and deletion flows where COPPA, GDPR Article 8 or local rules apply.
• Payments per market. Selling to individual learners in several countries means several gateways; PadhAI was designed for 9 behind one abstraction.
• Accessibility and exam-time load. Manual testing with assistive technology, and load tests shaped like a real exam hour.

Scope	Typical market range (2026)	What drives it
Basic learning app (MVP)	$40,000–$70,000+	Content delivery, learner profiles, quizzes, progress tracking
Mid-complexity platform	$70,000–$150,000+	Several roles (learner, teacher, parent, admin) or advanced workflows
Full-featured LMS	$150,000–$250,000+	Analytics, payments, video, third-party integrations, administration

Ranges from Cleveroad's June 2026 edtech app development guide, which does not price AI tutoring or LTI as separate lines. The same guide expects 3–5 months for a basic MVP; a narrower first version, such as one course or tutoring flow, can be shorter.

RAITHub's estimate separates the one-off build from the running costs you will carry, AI model spend and messaging fees included, and lists each channel and integration as its own line. LTI and institutional SSO are priced as new work, because they are not part of PadhAI's published description.

Why RAITHub for AI tutoring and learning platforms

Because RAITHub built PadhAI, a learning platform whose hardest problems, AI cost, answer correctness and reaching learners on the apps they already use, are the ones most tutoring products hit. RAITHub is a founder-led software studio, founded in 2024 in Dhaka, Bangladesh, working with clients worldwide.

• A multi-service platform, not a prototype. 11 services, 9 in Node/TypeScript and 2 in Python/FastAPI, so product logic and AI work each run in the language suited to them.
• AI cost designed in. The 70/20/10 model router sends each query to a lightweight, mid-tier or premium model by complexity.
• A tutor that checks itself. Socratic questioning, math verification and retrieval-augmented generation, with adaptive assessment driven by knowledge tracing.
• Built for emerging markets. Designed for 7 markets and 9 payment gateways, with the same experience on a PWA, WhatsApp and Telegram.
• Test discipline you can check. 750+ automated tests on TheSkinProof and 530+ on Sundor Skin, gated in CI.

RAITHub works on a fixed-scope build or a dedicated team, and you own the code through a present-assignment IP clause.

When RAITHub isn't the right fit

Hire someone else if you need proven learning outcomes from past work, course content production, formal certifications, or deep LMS integration experience on day one.

• You need your project run in a language other than English. RAITHub delivers in English: specs, demos and support. Your product can still serve learners in their own languages.
• You want published efficacy or usage results from past edtech work. RAITHub does not publish PadhAI user numbers or outcomes, so it cannot show you those.
• You need instructional design or course content. RAITHub's services are engineering: building the platform, not writing the curriculum.
• You need a vendor that can show shipped LTI integrations today. RAITHub's published work does not include one. A team that has already integrated with your target LMS will have fewer surprises.
• You need a SOC 2 or ISO 27001 certified vendor. RAITHub is not certified; see the security page.
• You want the lowest hourly rate with no tests. A CI-gated test suite is part of every RAITHub engagement.

How do you start an edtech project with RAITHub?

The EdTech service page sets out what RAITHub builds. To talk it through, book the free 15-minute technical audit. An NDA is signed before any detailed discussion.

It helps to describe your learners and their ages, the channels they use, the markets you serve and whether schools or individuals are paying. You will get a written audit memo either way.

Last reviewed: 28 September 2026.

Frequently asked questions

Has RAITHub built an edtech platform?

Yes. RAITHub built PadhAI, an AI-powered tutoring platform designed for 7 emerging markets, with adaptive assessment, a Socratic tutor and channel parity across PWA, WhatsApp and Telegram.

What is PadhAI's 70/20/10 model router?

It classifies each query by complexity and sends it to a lightweight, mid-tier or premium AI model, so simple questions do not pay premium-model prices. It is how PadhAI keeps AI cost per student viable.

Can RAITHub build an LMS?

RAITHub can engineer an LMS: courses, enrolment, assessment, progress tracking and payments, with a CI-gated test suite. LTI and institutional SSO would be new work on your project, estimated and tested as such.

What is knowledge tracing?

Knowledge tracing estimates, from a learner's sequence of answers, how likely they are to have mastered each skill. It lets a product choose the next question or practice task for that learner.

Does COPPA apply to my edtech app?

COPPA applies to operators of online services directed to children under 13, or with actual knowledge they collect personal information from a child under 13. Whether that includes you is a question for a privacy lawyer.

Is RAITHub's work WCAG certified?

No. RAITHub works to WCAG as a practice, including keyboard operation, contrast, captions and screen-reader testing, but it does not certify conformance. An independent auditor should produce any formal report.

How much does it cost to build an LMS or tutoring app in 2026?

Cleveroad's June 2026 guide puts a basic learning app MVP at $40,000–$70,000 or more, a mid-complexity platform at $70,000–$150,000 and a full-featured LMS at $150,000–$250,000 or more. These are market figures, and an AI tutor also carries running model costs. RAITHub gives a written fixed-scope estimate after a free 15-minute audit.

How long does it take to build a learning platform?

A focused first version, such as one course or tutoring flow with assessment and payments, typically takes 4–6 weeks at fixed scope. LTI, institutional SSO and multiple channels each add time and are estimated separately.