Hello, I'm
Shpetim Gashi
Data Scientist, Forecasting & Optimisation
Data science and software development at HELLA InnovationLAB from 2022 to 2026, from dual study to project lead of a reinforcement-learning research project. Python for time-series forecasting, linear optimisation and reinforcement learning, plus the tested software around the models: selected users worked productively with a material optimisation tool I further developed before I left.
Part-time M.Sc. in Data Science and Intelligent Analytics, weekend format, expected July 2027, attendable remotely and compatible with full-time employment. Available at short notice and open to relocation.
Featured Projects
Reinforcement Learning for Comfort Conditions in Buildings
Simulation-based reinforcement-learning experiments for control problems in building technology. PPO agents trained with stable-baselines3 and Gymnasium, Radiance-based simulation tooling for the environment, and TensorBoard for experiment tracking, with simulation assumptions and limitations kept explicit.
Manufacturing Optimisation
Material optimisation tool built on a colleague's linear-programming model in PuLP, which I further developed: extended the model, built its Streamlit and Flask interface and pytest tests, and built a CI pipeline that checks reference cases with known expected results before every release build. Selected users worked with it productively before I left.
Structured Comfort Feedback Platform
A university software-engineering project built around a real-world building-technology use case: a maintainable Flask application for collecting and managing structured user feedback, covering authentication and role handling, PostgreSQL persistence, administrative workflows, REST integration, automated testing, containerisation and CI/CD.
Demand Forecasting & Model Benchmarking
Time-series forecasting pipeline with the evaluation fixed before the models: LSTM models in TensorFlow/Keras benchmarked against (S)ARIMA baselines in statsmodels, with rolling evaluation windows and each horizon scored against the baseline.