LINKLEARN AI — NEURAL E-LEARNING
Deep Learning Course Recommendation Engine (Keras DLRM & Flask SVD on 3,522 Coursera Courses) + Claude-3-Opus Quiz Synthesis (/demo/linklearn-ai)
Master Thesis (UCBET // ENIC-NARIC Level 7)
AI Systems Researcher & Full-Stack Developer
2024
CodeIgniter 4 · Python Flask · Keras DLRM · Scikit-Surprise SVD
RMSE 0.8718 (-20.8% vs M1)
71.46% Recall@K · 70.53% Prec@K
3,522 Coursera × 610 Users
Claude-3-Opus + Flask SVD Demo
PROBLEM STATEMENT
Traditional e-learning portals suffer from information overload across thousands of courses and require instructors to manually transcribe video lectures and author assessments.
SYSTEM DESIGN
Preprocessed 3,522 Coursera courses with Python NLTK and merged them with MovieLens ratings (610 users × 9,724 items). Benchmarked Surprise SVD and two Keras DLRM architectures—proving that 50-dim User/Course Embeddings + Concatenate + Dense(100, ReLU) (Model 2, 526,901 params) cuts RMSE by 20.8% (0.8718 vs 1.1004)—and exposed inference via a Flask /recommend microservice consumed by CodeIgniter 4.
PRODUCTION EXECUTION
Delivered LinkLearn AI and its live interactive portfolio demo (/demo/linklearn-ai) featuring real-time Top-N Coursera recommendations, a Model 1 vs Model 2 telemetry benchmark, YouTubeTranscriptApi caption extraction + Claude-3-Opus (claude-3-opus-20240229) quiz generation, and Video.js Linked-Data timestamped annotations.
ENGINEERING & QA BLUEPRINT
Coursera + MovieLens ETL
3,522 Coursera courses (137.7+ KB) preprocessed with Python NLTK stopwords & merged with MovieLens 610×9,724 user-item ratings.
Keras DLRM & Flask SVD
50-dim User/Course Embeddings + Concatenate + Dense-100 ReLU (RMSE 0.8718, Recall@K 71.46%) & Flask /recommend SVD microservice.
Claude-3-Opus Quiz Synthesis
YouTubeTranscriptApi extracts lecture captions into MySQL and prompts claude-3-opus-20240229 (4000 tokens) to generate 10-question quizzes.
Video.js & Linked Data
Interactive Video.js player storing timestamped learner/instructor annotations and Semantic Web RDF/DBpedia URIs.
MySQL Schema: `users`, `courses` (3,522 Coursera entries), `ratings` (610×9,724 matrix), `videos_db` (YouTube ID + captions), `annotations` (timestamped notes), `assignments`.
GET `/demo/linklearn-ai` (Interactive Portfolio Demo) · Flask GET `http://localhost:5000/recommend?user_id={id}` (`svd_model.pkl`) · POST `/assignments/generate` (Claude-3-Opus).
CodeIgniter 4 MVC role separation (Student / Instructor), API key isolation for Anthropic Claude-3-Opus, CSRF protection, and sanitized TinyMCE/Video.js inputs.
Contract testing between CodeIgniter 4 `CourseController` and Python Flask `/recommend?user_id=...` JSON microservice, plus Anthropic Messages API payload verification.
Automated tests verifying `/demo/linklearn-ai` route rendering, course recommendation JSON structure, and MySQL database schema migrations.
20-epoch Keras training/validation MSE loss telemetry comparing Model 1 (Dot-Product: RMSE 1.1004) vs Model 2 (Concatenate + Dense-100 ReLU: RMSE 0.8718, -20.8% error).
// COMPLETE TECHNOLOGY STACK & TAGS
SCENE EXHIBITS
CLICK FRAME TO INSPECT IN LIGHTBOX