Scaling an EdTech LMS to 100,000 concurrent learners with AI-powered personalisation

Vidya Shala's founders saw a market underserved by every major EdTech player: working adults in Tier 2 and Tier 3 cities who needed vocational and professional upskilling but were excluded from existing platforms by language barriers (content was almost exclusively in English or Hindi), device limitations (low-end Android devices with inconsistent connectivity), and price sensitivity (₹10,000-₹40,000 course fees were out of reach). Their vision was a platform that delivered quality learning in Marathi and Kannada, optimised for 2G/3G connectivity, and priced at ₹499-₹1,999 per course.

Building an LMS is well-understood engineering. Building one that delivers video smoothly on a low-end Redmi device in a district town with variable connectivity is significantly harder. Adaptive bitrate video streaming, offline content downloads, and a mobile application that remained performant on 2-3 GB RAM devices were non-negotiable requirements. Any significant video buffering or app crashes would translate directly into learner drop-off in a segment where course completion rates on existing platforms were already below 15%.

Content personalisation — the core value proposition of AI-driven EdTech — is usually built on rich interaction data that a new platform does not have. Vidya Shala needed an adaptive learning system that could function meaningfully with limited historical data per user, using subject matter structure and learning science principles to personalise paths before enough behavioural data existed to drive pure ML recommendations.

We designed the platform for mobile-first delivery with progressive enhancement for desktop users. The video delivery pipeline used AWS CloudFront with HLS adaptive bitrate streaming, automatically adjusting quality from 240p to 1080p based on detected bandwidth. Content was structured into five-minute micro-modules rather than long lectures, significantly reducing the buffering risk on any given content unit. Offline download functionality allowed learners to download entire course modules on Wi-Fi for playback in offline mode — critical for learners with limited mobile data.

The adaptive learning engine was designed with a hybrid architecture: a knowledge graph representing each subject's prerequisite structure provided the initial personalisation layer before sufficient behavioural data existed per user. As learners completed assessments and modules, a collaborative filtering layer progressively increased its influence, identifying which learning paths taken by similar learners led to higher completion and assessment performance. The system adapted both the sequence of content and the type of content (video, reading, practice problem, worked example) based on individual performance signals.

Localisation was engineered into the data model from the start rather than retrofitted. All content metadata, UI strings, notification copy, and assessment text were stored with locale keys. A content management system built for Vidya Shala's in-house content team allowed course creators to publish in Marathi, Kannada, and Hindi from a single authoring environment, with language-aware search indexing ensuring that learners searching in their native language found relevant results.

The LMS launched with 120 courses across vocational skills, competitive exam preparation, and professional development, delivered in Marathi and Kannada. Video delivery on 3G connections achieved an average buffering rate of 0.8% — below the 2% threshold that research identifies as the point of measurable learner disengagement. The Android application maintained a 4.3-star rating on the Play Store across 15,000+ reviews, with positive mentions of performance on low-end devices appearing consistently in user comments.

The adaptive learning engine, fully operational at month six, demonstrated measurable impact by month nine: learners on adaptive paths showed a course completion rate of 41% versus 17% for learners on standard linear paths. Assessment scores at course completion were 23% higher for adaptive path learners. These metrics were validated through an A/B test where 30% of new enrolments were randomly assigned to the non-adaptive experience as a control group.

Infrastructure was designed for the scale ambition from the start. The platform handled 100,000 concurrent learners during a promotional campaign in month 18, serving video to all users without performance degradation. AWS auto-scaling policies adjusted compute capacity in response to demand spikes, and load testing prior to major promotional events became a standard pre-campaign practice. The platform now hosts 340 courses, 180,000 registered learners, and processes 2.1 million module completions per month.