An AI engine guiding learners to courses that match their goals and boost platform growth.
By: Itisha Dubey
Problem Statement & Objectives
Learners struggle with too many choices, leading to disengagement and churn. Our mission: transform a passive catalog into a proactive, intelligent learning guide.
+25%
User Engagement
Increase platform time via relevant discovery.
+30%
Completion Rate
Reduce dropouts with adaptive guidance.
+20%
Satisfaction Score
Deliver tailored, non-generic experiences.
Our Users
Serve four distinct user segments, each with unique needs.
Launches targeted campaigns (course bundles, nudges, live sessions)
Direct intervention before churn occurs
Features Recommended: Instructors & Creators
Content Gap & Demand Signal Dashboard
Analyzes learner search queries and forum threads
Integrates job market data
Shows exactly what content is missing and in-demand
AI Course Metadata & Quality Agent
Auto-generates skill tags and difficulty scores
Maps prerequisites from course content
Benchmarks quality against similar courses
Trend-Aware Content Brief Generator
Combines platform trends with industry signals
Generates ready-to-use content briefs
Includes module structure, target persona, and demand estimates
Non-Functional Requirements
Performance
Recommendations: <300ms response (95% of requests)
RAG Q&A: <2 seconds
Scalability
10x peak traffic support
50M+ events/day ingestion
Availability
99.9% uptime (learner-facing)
99.5% uptime (admin tools)
Compliance
GDPR & CCPA ready
EU AI Act transparency
Security & Data Retention
Encrypted at rest and in transit
Role-scoped access
Auto-purge after 24 months
Monitoring & Logging
Full recommendation context logged
Real-time drift alerting (15% threshold)
Extensibility
Modular architecture for new signals
Swappable LLM backends
Core System Architecture
Five modular layers enabling independent scalability and continuous learning.
Separation of recommendation engine from agent layer allows independent evolution of matching logic and proactive intelligence.
Trade-Off: Personalisation vs. Privacy
The Challenge: Better recommendations require more granular behavioral data. But aggressive tracking erodes trust and violates GDPR, CCPA, and EU AI Act requirements.
Our Decision: Privacy-First Personalisation
01
Tiered Data Collection
Low-sensitivity signals (completions, quiz scores) on by default
Fine-grained signals (video scrub positions) opt-in only
02
Early Anonymization
Compute dropout probability scores before sensitive data hits warehouse
Agents work with aggregated signals, not raw behavioral trails
03
Visible Explainability
Show learners why courses are recommended
"Recommended because you completed X and target Y role"
Builds trust through transparency
Bottom line: Slightly lower precision for opted-out users. Long-term gain: a platform learners trust.
Summary
Key Takeaways
1
Content-first
Prioritizes learner trust, content quality, and privacy.
2
Proactive, not passive
Agent layer acts as an intelligent companion, guiding and adapting in real-time.
3
Modular and extensible
Each layer is independently scalable and swappable (LLMs, data sources), built for long-term growth.
4
KPI-driven decisions
Engagement, completion, and satisfaction validated every feature prioritization and trade-off.
Assumptions & Future Roadmap
Key Assumptions
Data
Platform has sufficient learner interaction data for collaborative filtering
External market signals accessible via APIs
Users
Learners complete onboarding with honest, detailed input
Creators act on demand signals
Technology
LLM backend remains reliable and cost-effective
Infrastructure supports real-time event streaming
Compliance
GDPR & CCPA primary frameworks
Future markets addressed as needed
Future Enhancements
Multimodal Learner Profiling: Skill verification through assessments and credentials
Social & Cohort Learning: Peer-based recommendations and study groups
Employer Integration: Verified credentials recognized by hiring partners