Complete re-architecture of a conventional EdTech platform into a unified AI-native learning system serving K-12 learners across South Africa, Botswana, and India.
Executive Summary
A leading African EdTech platform engaged Cubet to rebuild its K-12 system from a conventional curriculum delivery application into a fully AI-native platform. The engagement covered complete re-architecture and AI implementation across data infrastructure, curriculum engine, assessment logic, content processing, learner interaction, and analytics.
The rebuilt platform operates as a unified AI system. Personalisation, curriculum adaptation, assessment generation, predictive intervention, and conversational learning run from a shared real-time learner profile model. Every capability is integrated, not modular.
Outcomes measured across equivalent learner cohorts pre and post transformation:
55% increase in learner engagement. 45% improvement in assessment scores. 2x faster course completion. 4x platform usage growth. 2.5x improvement in course completion rates. 3x improvement in quality of learning score.
Industry: EdTech / K-12 Geography: South Africa · Botswana · India Engagement: Full platform re-architecture and AI transformation Services
- AI Personalised Learning
- Adaptive Curriculum Engine
- Intelligent Assessment
- Conversational AI Agent
- Learning Analytics
- NLP Content Processing
- Data Engineering and Pipelining
- Custom Generative AI Applications
Platform Context
The client is one of Africa's leading K-12 EdTech platforms, with significant deployment in rural schools across South Africa and Botswana where access to individualised instruction has been constrained by geography and resource availability. The platform operates across multiple curriculum frameworks serving learners across all K-12 grade levels in three countries.
The pre-transformation platform delivered structured curriculum, progress tracking, and teacher reporting. Its architecture reflected conventional EdTech design assumptions: a relational data model optimised for storage and linear progress tracking, rule-based curriculum sequencing, static scheduled assessments, and no real-time learner profile model.
Why a Full Rebuild
The limitations were architectural, not functional. The existing system performed its designed operations reliably. It could not support AI capabilities that depend on infrastructure it was not designed to provide.
Real-time personalisation requires continuous signal ingestion and millisecond-level inference availability. The existing data layer was batch-oriented. Adaptive curriculum sequencing requires a graph-based content model with dynamically weighted relationships. The existing curriculum was a fixed linear content tree. Generative assessment requires a semantic content graph and a live learner state model. Neither existed. A conversational agent grounded in learner context requires a continuously updated profile model accessible at inference time. The architecture had no such model.
Retrofitting these capabilities would have introduced compounding technical debt: inference pipelines disconnected from the core data model, latency failures under real-time load, and personalisation operating on stale signals. The decision was to rebuild.
Architecture Overview
The rebuilt platform is structured around a centralised real-time learner profile model serving as the shared intelligence layer for all AI capabilities. Every component reads from and writes to this model continuously, creating a closed-loop system in which learning signals, curriculum decisions, assessment outcomes, and interaction history are unified.
Data infrastructure is built on a real-time event streaming architecture with millisecond-level signal availability. The curriculum layer is an adaptive concept graph. The assessment layer is generative, connected to the semantic content graph and live learner state. The conversational agent operates via a RAG layer connected to the learner profile, content graph, and real-time session context. The analytics layer draws from the same unified data model.
AI Capabilities
Real-Time Learner Profile Engine Every learner interaction generates events ingested in real time to update a centralised profile model continuously. The profile captures demonstrated mastery by concept, struggle indicators, learning velocity, engagement depth, behavioural patterns, and where clinically relevant, learning disability markers that govern downstream AI behaviour across all capabilities. Profile data is queryable at millisecond-level latency for real-time inference.
AI Personalised Learning Paths A multi-signal scoring model evaluates the learner profile continuously to determine the optimal next content unit at each moment. Input signals include concept mastery scores, struggle indicators, learning velocity relative to cohort baseline, engagement depth, and time-since-reinforcement. The model re-evaluates after every interaction. No two learners follow the same sequence. Pacing is determined by demonstrated readiness, not a fixed schedule.
Adaptive Curriculum Engine The curriculum is modelled as a directed concept graph with nodes representing learning concepts and edges representing prerequisite and reinforcement relationships. Edge weights are initialised from pedagogical design and updated dynamically from cohort performance data. Weak mastery increases reinforcement edge weight, surfacing prerequisite content before progression. Strong mastery reduces reinforcement weight and accelerates progression. The graph updates continuously without manual reconfiguration.
Predictive Learning Layer Predictive models trained on anonymised cohort data identify early signals preceding disengagement, knowledge gap formation, and learning velocity decline. Models operate on rolling windows of learner profile data, generating probability scores at configurable time horizons. When scores exceed defined thresholds, intervention flags surface to teacher dashboards before performance drops become visible in formal assessments.
Intelligent Assessment Engine Assessment items are generated contextually from the semantic content graph and the learner's current profile state, not retrieved from a static question bank. Item difficulty, framing, and concept focus are calibrated to the learner's demonstrated level at the moment of assessment. Results write back to the learner profile immediately, updating mastery scores and feeding the next iteration of the personalisation engine. Assessment is a continuous feedback mechanism, not a periodic evaluation event.
NLP Driven Content Processing All curriculum content is processed through an NLP pipeline that extracts semantic structure and indexes content by concept, complexity, prerequisite relationships, and pedagogical intent, producing a semantic content graph. Content is mapped to learner needs at query time based on semantic relevance to the current learner profile state, not keyword proximity.
The Conversational Learning Agent Built on a fine-tuned large language model with a RAG layer connecting the agent to three data sources at inference time: the learner's live profile model, the semantic content graph scoped to the learner's current curriculum position, and the real-time session context.
Every response is grounded in the intersection of learner profile, curriculum context, and session state. The agent handles concept clarification, alternative explanations, questions about prior content, and extended exploratory conversations. Context is maintained across sessions via the persistent profile model.
For learners with identified learning disabilities, behavioural markers in the profile model govern the agent's communication approach, pacing, and explanatory style without requiring self-identification at the interaction level.
The agent operates across three role-governed stakeholder modes.
Learner mode: full conversational capability grounded in individual profile and curriculum context.
Teacher mode: class-level performance intelligence, individual intervention recommendations, concept difficulty signals aggregated across the cohort.
Parent mode: profile-grounded progress summaries, proactive support flags, and responses to questions about learning trajectory.
One architecture. Three interaction modes. All responses governed by role permissions and the unified learner data model.
Learning Analytics and Insights The analytics layer draws from the unified profile model, providing consistent data across all stakeholder views. The platform-generated quality of learning score aggregates mastery signals, retention indicators, and concept application performance into a continuous real-time metric tracking genuine knowledge acquisition rather than content consumption. Updated continuously, not at assessment cycle boundaries.
Outcomes
| Metric | Measurement Basis | Result |
| Learner engagement | Active learning time, session frequency, content interaction depth | 55% increase |
| Assessment scores | Average performance, equivalent curriculum and cohorts, pre vs post | 45% improvement |
| Course completion speed | Time to complete equivalent curriculum content | 2x faster |
| Platform usage | Total active usage, transformation period vs equivalent pre-transformation period | 4x growth |
| Course completion rates | Completion percentage, enrolled learners, equivalent cohorts | 2.5x improvement |
| Quality of learning score | Platform-generated, knowledge retention and concept mastery | 3x improvement |
The quality of learning score is platform-generated, aggregating demonstrated mastery, retention signals from spaced assessment, and concept application performance. A learner completing content without retained understanding scores lower than one demonstrating concept application in subsequent assessments.
Implementation Context
The engagement required parallel workstreams across real-time data infrastructure, AI system design, NLP pipeline development, LLM fine-tuning, RAG architecture, curriculum graph modelling, and product engineering, integrated continuously across the shared data layer.
The client's education specialists were embedded throughout, ensuring AI system behaviour reflected pedagogical intent. Curriculum graph structure, concept relationships, mastery thresholds, and intervention trigger logic were defined collaboratively between Cubet's engineering team and the client's domain specialists.
Deployment Significance
The platform serves a substantial learner population in rural communities across South Africa and Botswana where differentiated instruction and individual tutoring support have been structurally unavailable. The 2.5x improvement in course completion rates in this context represents a material change in educational outcomes for a population with limited alternative access to equivalent quality of instruction.
The unified real-time learner profile model and its integration across all AI capabilities is the technical architecture that makes genuinely individualised learning achievable at the scale and in the deployment environments the platform operates in.
To discuss a similar engagement: Contact Cubet's AI engineering team
Get in touch
Kickstart your project
with a free discovery session
Describe your idea, we explore, advise, and provide a detailed plan.
