AI-Driven Student Success Forecaster
Hackathon project (Udacity Palestine Launchpad). Predicts student outcomes and flags at-risk students with an explainable ML pipeline and a Flask web app.
Overview.
Built during the Udacity Palestine Launchpad Hackathon, our team delivered an AI web app that forecasts student performance and surfaces early-warning insights for educators. We placed 6th overall with one of the highest accuracies on the Kaggle leaderboard used for judging. The app is deployable and accessible via a simple browser UI.
Key features
- End-to-end pipeline: data cleaning, feature engineering, train/validation split, and model selection with cross-validation.
- Explainability: feature importance and per-prediction rationale to support interventions.
- Web app: lightweight Flask interface for single-student form inputs and batch CSV scoring.
- Production artifacts: serialized model and preprocessing pipeline for reproducible results.
- Team delivery: clear repo structure, demo screenshots, and setup instructions for evaluators.
Tech stack
- Python
- Pandas
- NumPy
- scikit-learn
- Flask
- Jupyter
- joblib
- Matplotlib
Architecture (simplified)
- Ingest & clean: validate ranges, impute missing values, encode categoricals, scale numerics.
- Modeling: try several scikit-learn baselines; pick the best via cross-validation (RMSE/MAE/Accuracy depending on task).
- Export: persist the fitted model + preprocessing with joblib.
- Serve: Flask app loads artifacts at startup and exposes predict endpoints plus a browser form.
- Evaluate: track leaderboard score (Kaggle) and compare iterations to improve generalization.
Dataset
Tabular student performance data provided for the hackathon (features such as demographics/attendance/engagement where available). The code expects a CSV schema matching the training set and validates inputs before scoring.
Links
- Code repository: Students_grade_forcasting (GitHub)
- Live demo: Student Success Forecaster Web App
- Udacity: Udacity · Kaggle: Kaggle