L

lightgbm

A production-grade machine learning system for classifying celestial objects — stars, galaxies, and quasars — from 500,000 photometric observations sourced from the Sloan Digital Sky Survey (SDSS DR18). The pipeline covers LOF-based outlier detection, SMOTEENN class balancing, and SelectKBest feature selection, with six classifier configurations benchmarked against each other. Random Forest achieved the highest accuracy at ~99.51%, while LightGBM was selected for deployment due to its faster inference, smaller footprint, and clean ONNX export path. The system is fully containerized with Docker, backed by a CI/CD pipeline, and served live via a Gradio interface on Hugging Face Spaces.

Self-hosted forecasting + prediction service. Five zero-shot time-series foundation models (Chronos-2, TimesFM 2.5, Moirai-2, Toto-1, Sundial) across six forecast types, plus nine supervised tabular ML backends (LightGBM, XGBoost, sklearn family) with calibrated / stacking / diversified meta-learners. Unified REST API + MCP server.

Fundamental theory and practice in Data Science (DS). 🧮

This project uses LightGBM to make associations between cladograms and its terminals' metadata, generalizing the idea first presented by Vieira Alves et al. (2025, DOI: 10.1111/cla.12610).

Final project of course ML junior by Skillbox. Goal: Build model to predict client's loan default. Metrics: ROC-AUC

This is a test task for Longevity InTime AI powered company.