Adaptive Learning System | David Mike-Ewewie

The Challenge

Course Recommendation Crisis

Engineering students receive poor course recommendations 66% of the time, leading to extended graduation times, increased dropout rates, and unnecessary financial burden on students.

ABET Accreditation Constraints

Any recommendation system must strictly adhere to ABET accreditation requirements, including prerequisite chains, core requirements, and credit hour minimums - making this a constrained optimization problem.

Scale & Performance

System must handle 32,000+ students making course selections simultaneously during registration periods, requiring sub-40ms response times to prevent registration bottlenecks.

The Process & My Contribution

Led the design and implementation of a production-ready adaptive learning system using advanced reinforcement learning techniques.

1

Problem Formulation

Reformulated course recommendation as a contextual multi-armed bandit problem, where each course sequence is an "arm" and student features form the context. This enabled real-time personalization while maintaining exploration-exploitation balance.

2

Algorithm Development

Implemented Thompson Sampling with Beta-Bernoulli conjugate priors for efficient Bayesian inference. Developed custom reward functions incorporating GPA outcomes, time-to-graduation, and ABET compliance.

3

Constraint Satisfaction

Built constraint satisfaction layer ensuring all recommendations meet ABET requirements. Implemented efficient graph algorithms for prerequisite validation and credit hour optimization.

4

Production Deployment

Engineered horizontally scalable API supporting 520 requests/second with 38ms p99 latency. Implemented A/B testing framework for safe rollout and continuous improvement.

Technical Architecture

Contextual Bandit Framework

The system uses a sophisticated contextual multi-armed bandit approach that personalizes recommendations based on student features while continuously learning from outcomes.

Thompson Sampling Implementation

1. Context Extraction: Student features including GPA, completed courses, major, learning style preferences

2. Posterior Sampling: Sample from Beta distributions for each course sequence based on historical success rates

3. Constraint Filtering: Remove sequences violating ABET requirements or prerequisite chains

4. Action Selection: Choose sequence with highest sampled reward, ensuring exploration

5. Outcome Tracking: Update posteriors based on student performance (GPA, completion rate)

Performance Optimizations

Caching Strategy

Multi-level caching with Redis for frequently accessed course sequences and student profiles, reducing database load by 78%.

Batch Processing

Asynchronous batch updates for posterior distributions during off-peak hours, maintaining real-time performance during registration.

Feature Engineering

Dimensionality reduction using PCA on student features, reducing computation time by 45% without accuracy loss.

Tech Stack

Core Algorithm

Python NumPy SciPy Scikit-learn

API & Backend

FastAPI PostgreSQL Redis Celery

Infrastructure

Docker Kubernetes AWS ECS CloudWatch

Monitoring

Prometheus Grafana MLflow Datadog

Results & Impact

66%

Reduction in Poor Recommendations

32K+

Students Served

520

Requests/Second

38ms

P99 Latency

Experimental Results

Cumulative Regret Analysis

Thompson Sampling achieved 34% lower cumulative regret compared to ε-greedy and 52% lower than random selection over 10,000 iterations.

Student Outcomes

Students following recommended sequences showed 0.3 higher average GPA and 1.2 semesters faster graduation compared to control group.

Engagement Metrics

92% of students reported satisfaction with personalized recommendations, with 78% following suggested course sequences.

System Reliability

Maintained 99.97% uptime during critical registration periods, with zero ABET compliance violations in 500,000+ recommendations.

Publication Details

WIP: Data-Driven Adaptive Curriculum — Personalizing Academic Pathways for Enhanced Engineering Student Success

D. Mike-Ewewie, P. Lim, A. Sotelo, P. Kumar

IEEE Frontiers in Education Conference (FIE) 2025 - ACCEPTED

Abstract

This work-in-progress paper presents a novel application of contextual multi-armed bandits to personalize course recommendations for engineering students. By formulating curriculum planning as an online learning problem, our system adapts to individual student needs while maintaining strict adherence to ABET accreditation requirements. Preliminary results demonstrate a 66% reduction in suboptimal course recommendations, with the production system successfully handling 32,000+ students at 520 requests/second with 38ms latency. The Thompson Sampling approach significantly outperforms traditional advising methods, showing promise for improving student outcomes at scale.

Key Innovations

Novel Problem Formulation: First application of contextual bandits to curriculum recommendation with hard constraints
Efficient Constraint Handling: Custom graph algorithms ensuring ABET compliance with O(1) validation time
Production-Ready Scale: One of the few academic ML systems deployed at production scale with published performance metrics
Interpretable Recommendations: Explainable AI features showing why specific courses were recommended
Continuous Learning: Online learning approach that improves with each semester's outcomes

Future Work

The accepted paper opens several avenues for future research and development:

Extend to multi-objective optimization including career outcomes and skill development
Implement deep contextual bandits using neural networks for richer feature representations
Develop transfer learning approaches for cross-university deployment
Create real-time dashboards for advisors to monitor and intervene when needed
Integrate with university LMS systems for seamless deployment

Adaptive Learning System (FIE 2025)