Applied AI
Research Archive

I approach machine learning as a full research-to-production pipeline. I start with Exploratory Data Analysis (EDA), data cleaning, preprocessing, feature engineering, and visualization to understand the real structure of the data before choosing a model.

From there, I build a strong baseline first, then improve it through cross-validation, metric tracking, error analysis, threshold calibration, and controlled experiments. I care about what each model learns, where it fails, and whether the final result is actually useful beyond a notebook score.

My current project archive now covers medical AI, computer vision, NLP, tabular machine learning, regression, classification, object detection, segmentation, web scraping, and data engineering. The newer projects go beyond notebooks into deployed systems with APIs, typed frontends, databases, Docker, migrations, encrypted persistence, automated tests, and CI/CD.

For deep learning projects, I focus on practical architecture decisions: CNN backbones for image classification, YOLO-style detection for localization, U-Net-style segmentation for medical masks, and sequence models such as BiLSTM, attention, Siamese networks, embeddings, TF-IDF, and NLP preprocessing for text tasks.

For classical machine learning, I still treat algorithms as serious production tools. I compare linear models, SVMs, tree models, Random Forests, XGBoost, LightGBM, CatBoost, voting ensembles, feature engineering strategies, and tabular validation workflows before deciding whether deep learning is really necessary.

The direction is moving from isolated experiments toward complete AI systems: Lung Disease Detection, Quora Question Pairs, Disaster Tweets, SMS Spam, Global Wheat Detection, Cassava Leaf Diseases, TMDB Revenue, and earlier classical ML/data-analysis projects all connect model performance with real implementation practice.

On the application side, I connect trained models with real interfaces using FastAPI, React and TypeScript, Docker and Nginx, SQL databases, Alembic migrations, ONNX Runtime, encrypted prediction logging, and GitHub Actions CI/CD. Separate projects explore MLflow, Airflow, and Kafka around the model lifecycle, while Optuna supports controlled hyperparameter tuning. The goal is not only to train models, but to understand how they become reliable AI products.