Personalised Learning at Scale: How PurpleGene® is Redefining K-12 Education in India

India's school system is the second largest in the world by enrolment — 260 million children across 1.5 million schools. Yet the Annual Status of Education Report (ASER) consistently shows that over 50% of Class 5 students cannot read a Class 2-level text, and nearly 70% cannot perform basic division. These are not outliers. They are systemic outcomes of a curriculum designed for the median student that fails both the struggling learner who falls further behind each year and the advanced learner who disengages from material they mastered months ago.

Personalised learning has long been the proposed solution. But personalising education at the scale of 260 million children is an engineering challenge of extraordinary complexity — particularly when the infrastructure available is a shared Android tablet and an intermittent 2G connection.

PurpleGene®'s core is what we call the Adaptive Mastery Engine (AME) — a real-time recommendation system that continuously models each student's knowledge state across 1,400 curriculum nodes mapped to the NCERT framework.

The AME does not simply track right and wrong answers. It models the process: how long a student spent on a question, whether they used the hint system, whether they made the same error type in a different context, and how their accuracy changes across time-of-day sessions. From this signal, the engine infers both current knowledge state and learning velocity — how quickly the student is likely to consolidate a new concept — and selects the next learning experience accordingly.

We use a hybrid approach: a knowledge graph-based Bayesian Knowledge Tracing model for concept mastery estimation, combined with a collaborative filtering layer that draws on patterns from 1.2 million learner histories to surface content types that work well for students with similar profiles. The system retrains nightly on anonymised, aggregated interaction data.

The first version of PurpleGene® required a stable internet connection. In urban private schools, it worked beautifully. In government schools in rural Bihar or tribal districts of Odisha, it was unusable. Students would lose session progress mid-lesson, video content would buffer indefinitely, and teachers would abandon the platform after a week of frustration.

We rebuilt the architecture from scratch around an offline-first model. The entire curriculum graph — 14,000 learning objects including videos, interactive exercises, and assessments — is pre-cached on the device during any available connectivity window. The AME runs entirely on-device using a compressed inference model (our custom-quantised variant of a small language model, under 80MB). Student progress syncs to the server in delta packets when connectivity is available.

The result: PurpleGene® is fully functional with zero connectivity. It degrades gracefully to 2G for sync. And it never loses a student's learning record.

In a randomised controlled trial conducted across 240 government primary schools in Rajasthan over 18 months, students using PurpleGene® for 45 minutes per day showed:

— 34% improvement in foundational literacy scores versus control group — 41% improvement in numeracy outcomes — 28% reduction in dropout rates among students in the bottom quartile of initial assessment

Teacher satisfaction scores — a critical factor in EdTech adoption — averaged 4.3 out of 5. Teachers reported that the platform's dashboard, which surfaces each student's current mastery level and recommended next steps, reduced their lesson preparation time by an average of 40 minutes per day.

The most surprising finding: students in the PurpleGene® cohort spent an average of 22 additional voluntary minutes per day on the platform beyond their scheduled sessions. In populations where screen fatigue and disengagement are persistent challenges, this intrinsic motivation signal is the strongest evidence we have that the personalisation is working.