Independent projects, guided builds, academic research and selected professional contributions. Source, demos and limitations are included where they can be shared.
9 projects
Swiss Energy Forecast
Built a pipeline that combines electricity and weather data to produce 48-hour Swiss energy forecasts with prediction intervals.
Load electricity and weather feeds into PostgreSQL, build time-aligned features, train LightGBM models, and track runs with MLflow. FastAPI serves forecasts locally; the public React dashboard reads scheduled JSON exports.
Validation and limits
Tests cover feature alignment between training and inference. The dashboard shows a historical backtest. Split conformal calibration targets 90% coverage; it does not guarantee coverage on future operational data.
Swiss BGE RAG
Built a search and question-answering prototype for Swiss Federal Supreme Court rulings, with cited answers and a fallback when retrieved passages don’t support a response.
Independent project · Prototype
Data Science & Applied AI
Python · Qdrant · BM25 · Streamlit · RAGAS
Project details
Implementation
Split rulings by section, combine dense embeddings and BM25 in Qdrant, and rerank passages before prompting the model to cite case numbers.
Evaluation and limits
The repository reports RAGAS results on 26 questions, with stronger context retrieval scores than answer faithfulness or relevance. The demo indexes a small corpus and does not independently verify citations against the source.
SQL Data Warehouse
Adapted a guided SQL Server warehouse project to PostgreSQL, turning CRM and ERP CSV files into reporting views.
Guided adaptation · Prototype
Data Engineering · Analytics & Analytics Engineering
PostgreSQL · SQL
Project details
Implementation
Load source files into Bronze tables, use PostgreSQL procedures to clean and standardize them in Silver, and expose Gold fact and dimension views as a star schema.
Validation and limits
The repository includes SQL quality checks and architecture documentation. This is a tutorial-derived build using example data.
I build Python document-processing and data workflows, automate recurring tasks, and contribute to Power BI reporting in my Working Student QEHS role.
Professional work · Ongoing work
Process Automation · Analytics & Analytics Engineering
Python · Power BI · Power Automate · Power Apps
Project details
My contribution
Developed document processing to extract, validate and structure shipment information. My work also includes transport-data processing for emissions estimates and Power Platform tools for recurring operational tasks.
Sharing scope
This overview describes my contributions. Internal workflows, code, data and screenshots are not shared.
Smart City Data Engineering
A streaming prototype that generates simulated vehicle telemetry, transforms it in Spark, and configures Parquet output to Amazon S3.
Independent project · Prototype
Data Engineering
Kafka · PySpark · Amazon S3 · Parquet · Docker
Project details
Implementation
A Python producer publishes synthetic signals to Kafka. Spark Structured Streaming processes the events; the repository defines S3 sinks and checkpoints and supplies a Docker Compose environment.
Limits
This is a simulation. The source and configuration do not establish production deployment or throughput.
Evaluate Student Summaries
An HSLU team project combining readability features, neural models and a transformer to score student summaries for content and wording.
Academic team project · Prototype
Data Science & Applied AI
Python · LightGBM · DeBERTa · PyTorch
Project details
Team scope
The repository documents the team’s three-path ensemble and experiment notebooks; it does not separate individual implementation contributions.
Tennis Analysis
A computer-vision prototype detecting players, the ball, and court keypoints to project positions onto a mini-court.
Independent project · Prototype
Data Science & Applied AI
Python · YOLO · ResNet50 · OpenCV
Project details
Implementation
Use separate YOLO trackers for the players and the ball, interpolate gaps in the ball track, and map detections onto a scaled court using ResNet50 keypoints.
Automated Assessment of Linguistic Abilities of Language Models
Extended the Holmes evaluation framework with an interactive probing interface and compared quantized and full-precision language models. Thesis source and detailed results are confidential.