Muhammed Mustafa KAPICI

Developed a full unsupervised learning pipeline for credit-card customer segmentation. Includes preprocessing (imputation, winsorization, scaling), engineered behavioral features, dimensionality reduction (PCA, UMAP), and clustering with K-Means, GMM and HDBSCAN. Integrated model evaluation (silhouette, DB index, ARI) and built a Streamlit dashboard for interactive exploration.

Built a production-style PyTorch training framework with MLflow experiment tracking and Model Registry. Added a FastAPI inference service with Prometheus metrics, Dockerized workflows, Makefile automation and dual frontends (Streamlit + Gradio). Demonstrates a complete MLOps-ready lifecycle from training to deployment.

Architected a high-performance regression pipeline using XGBoost for real-world price prediction, achieving a 0.94 R2 score and a low MAE of 1.88. Features preprocessing, domain-specific feature engineering, hyperparameter tuning, model explainability, versioned model management and deployable artifacts.

Developed a full-stack Retrieval-Augmented Generation system using LangGraph, Google Gemini and ChromaDB. Includes a modular document ingestion pipeline, embedding + retrieval workflow, streaming-enabled FastAPI backend, a React/Vite frontend, and Docker-Compose orchestration. Demonstrates modern LLM engineering and production-grade RAG design.

Created a multi-model classification pipeline for clinical heart-disease prediction. Covers preprocessing, feature engineering, stratified cross-validation, hyperparameter tuning and probability calibration. Combines neural networks (MLP) and tree-based models (XGBoost) to analyze structured medical data.

Built a high-frequency time-series forecasting engine for crypto markets using LSTM, GRU and Transformer architectures. Includes automated market data ingestion, technical-indicator generation, volatility analysis, S2S modeling, hyperparameter optimization and full backtesting simulations tailored for financial prediction tasks.

Designed an end-to-end satellite image classifier using state-of-the-art architectures (Swin Transformer, ConvNeXt). Features mixed precision, warm-up cosine LR scheduling, weighted sampling/Focal Loss for imbalance, patch-based inference, TTA, and a deployable Gradio UI. Demonstrates advanced CV techniques on remote sensing data.

Engineered a deep-learning pipeline for astronomical PSF detection using a customized ResNet backbone. Workflow includes simulation-driven augmentation (blur, atmospheric distortion), domain-adapted preprocessing, fine-tuning, and morphological post-processing. Shows capability in bridging astrophysics and convolutional networks.

Created a lightweight, pip-installable library that simplifies ML workflows: preprocessing utilities, classification/regression/clustering models, automated evaluation metrics, and integrated plotting. Designed for rapid prototyping and educational or experimental ML tasks.