PhD in Industrial Management (Operations Research). Founder & Lead AI Consultant at Homat AI, where I build data mining pipelines, ML/DL systems, and AI agents for real businesses.
This is a portfolio of end-to-end applied ML projects — each one built with a real dataset (or a transparently-labeled synthetic one when live data wasn't reachable), a rigorous evaluation methodology, and an honest discussion of what the numbers actually mean.
Connect: LinkedIn · Kaggle · homatai.com
| Project | Category | What it shows |
|---|---|---|
| Crypto Market Volatility Analysis | EDA | BTC/ETH/SOL/ADA volatility, drawdowns, and cross-asset return correlation on live CoinGecko data |
| Stock Market Seasonality & Anomaly Detection | EDA | Calendar effects (Monday/turn-of-month/January) + z-score & Isolation Forest anomaly detection across 5 sectors |
| Credit Scoring — Explainable AI | Machine Learning | XGBoost default-risk model with global & per-applicant SHAP explanations over 10K loan applicants |
| Real-Time Fraud Detection | Machine Learning | XGBoost fraud classifier, time-based split (not random!), precision/recall trade-offs, latency benchmark |
| ML-Based Portfolio Optimization | Machine Learning | Markowitz optimization with Ledoit-Wolf shrinkage, efficient frontier, risk parity, walk-forward backtest |
| Enterprise RAG Knowledge Base | LLM / NLP | Hybrid BM25 + LSA retrieval via Reciprocal Rank Fusion, evaluated with Recall@k / MRR@k against labeled questions |
| Financial News Sentiment Analyzer | LLM / NLP | TF-IDF + Logistic Regression sentiment classifier — exposes an evaluation-leakage gap between random and held-out-template splits |
| Interactive Knowledge Retrieval Agent | LLM / Agents | Tool-routing agent (TF-IDF router + memory) evaluated on single-turn routing accuracy and multi-turn coreference resolution |
| Industrial Defect Detection on Assembly Lines | Computer Vision | HOG + LBP features with Random Forest / SVM, per-defect-subtype accuracy breakdown (not just overall accuracy) |
| Real-Time Helmet and Safety Gear Detection | Computer Vision | HSV color segmentation + contour heuristics, evaluated at three levels: detection, localization (IoU), compliance decision |
All 10 projects in this portfolio are complete and published (GitHub + Kaggle + write-ups). Each one includes an honest limitations section — the point isn't a leaderboard number, it's showing the evaluation methodology and where it breaks.
Every project here follows the same standard: a clearly documented data source, a methodology chosen to avoid the evaluation mistakes that make results look better than they are (leakage, random splits on time series or templated text, cherry-picked metrics), and a README that states the limitations plainly. If a result looks too good, the project says why — and shows the honest number next to it.