Academic project — Data & Machine Learning
Depression Project
Flask/MongoDB application, semantic search and analyses of 19,999 records.
01
The starting point
Combine data management, semantic search and statistical exploration on a depression-related dataset. The project combines application development with critical model evaluation.
Academic project combining a Flask/MongoDB Atlas data management application with separate analysis scripts. Exploration, K-means, PCA, classification and regression are complemented by MiniLM semantic search. This is an analytical exercise, not a clinical diagnostic tool.
19 999Analyzed records
384Embedding dimensions
4K-means clusters
02
How it works
- 1MongoDB Atlas data
- 2Flask application · CRUD
- 3MiniLM embeddings
- 4Semantic search
Architecture & implementation
- Flask application with CRUD operations on MongoDB Atlas.
- Semantic search using MiniLM embeddings.
- Separate analytical scripts: exploration, K-means, PCA, classification and regression.
Technical decisions
- Flask for a lightweight architecture suited to a web analytical project.
- MongoDB Atlas for storage flexibility and easy pipeline integration.
- Clear separation between data management, statistical analysis, modeling, and visualization.
- Semantic search to go beyond a simple data viewing interface.
Lessons learned
- A relevant data project requires as much rigor in data management as in modeling.
- Visualization plays a key role in making results interpretable.
- An analytical web interface makes a scientific project far more usable.
Possible next steps
- Improve the UI to make exploration more intuitive.
- Strengthen comparison between machine learning models.
- Further document pipelines and analytical assumptions.
03
Results & practical value
- An application supports browsing and managing records; embeddings add search based on text meaning.
- Separate scripts explore groups using K-means/PCA and evaluate classification and regression tasks.
- Comparison with a simple baseline is part of the result: the sleep model does not improve accuracy over the majority baseline.
Delivered so far
- Functional data management application and semantic search.
- Analyses of 19,999 records in separate scripts.
- Model performance compared with baselines and limitations discussed.
04
Scope & limitations
- Predictive analyses are separate scripts, not features all integrated into the web application.
- This project does not diagnose depression or validate clinical use.
Source: 2025 data warehousing project code and report. Diagram shows the application workflow; statistical analyses run separately.