Samsung Innovation Campus capstone
Dec. 2025 – Jan. 2026
Results
- Engagement model: test R² 0.703, MAE 0.314 on 1,963 held-out lessons
- Bac-content filter: precision 0.942, recall 0.821 on 282 hand-labelled videos
Links
Much of the revision for the Algerian Baccalaureate now happens on YouTube. For our Samsung Innovation Campus capstone, the four of us set out to understand which Bac lessons engage students, and why.
What I built
I built the data side of the pipeline. I collected the lessons through the YouTube Data API, then built the filter that keeps only Bac content: channel priors combined with TF-IDF term discovery. Against 282 videos we labelled by hand, it reaches a precision of 0.942 and a recall of 0.821. On top of that I engineered the engagement features and built the predictive model: test R² 0.703 and MAE 0.314 on 1,963 held-out lessons.
I also proposed extracting best practices from the collected data. A teammate built that into the RAG-based recommendation agent over a local knowledge base, and another teammate handled the transcription step.
After the capstone
Our team won first place. Afterwards I prepared the project for an open-source release as BacStage: the features are now fitted on the training split only, with tests, a documentation site and a demo. The public repository holds the code only; the collected data and the trained models stay private.