// work
Projects
From fine-tuning transformers to live trading systems — everything I've shipped.
Three-way benchmark comparing English DistilBERT (zero-shot), mBERT (multilingual), and BERTurk on Turkish sentiment analysis. Proves language-specialized models decisively outperform multilingual ones — BERTurk's morphological awareness handles agglutinative structures that general tokenizers can't. Interactive live demo lets you test all three models in real time.
Fine-tuned DistilBERT on the IBM Argument Quality dataset to classify debate arguments as PRO or CON. Explores stance ambiguity — how identical statements can support opposing positions depending on topic context. Trained on 30,000 annotated examples with a perfectly balanced dataset.
Fine-tuned DistilBERT on WikiANN English (20k sentences) for PER / ORG / LOC entity detection. Manually reviewed 80k test tokens — found 74% of errors were misclassifications rather than missed detections, mostly LOC/ORG confusion on ambiguous geopolitical names. Some model predictions outperformed original labels.
Professional TradingView indicator suite for confluence zone detection. Auto-detects the trading mode (intraday / hourly / macro) and loads the relevant levels — ARMAS bands, VWAP, FVGs, market profile, open interest signal matrix, CVD divergence, session tracking, and a full scoring engine that ranks zones by confluence strength.
Hybrid RL agent (MaskablePPO) learning trade timing for the ARMAS mean-reversion strategy on SP500 futures. Iterated through 6 reward designs to escape degenerate policies. Key finding: signal edge (+0.056% per touch) is smaller than commission friction (~0.2%), making ML application premature. Documented as a rigorous negative result.
Mean-reversion Expert Advisor for MetaTrader 5 evolved over 18 iterations from a Python backtest to a live production system. Uses EWM bands with RSI, ATR, and slope filters. Optuna for hyperparameter optimization and ML training data generated from live trade history.
Full-featured trading journal built with Python and Streamlit. Multi-journal support, MT5 report import, pre/post/position analysis with screenshot upload, equity curve, calendar heatmap, analytics by session/hour/day, Hall of Fame, and PDF export. Used daily in real trading.
AI-powered exam prep tool for the Turkish ALES university entrance exam. Pulls from a database of 1,400+ real past questions to generate personalized practice sets and step-by-step solutions using Gemini 2.5 Flash. Includes study formula cards ranked by historical frequency.
Emergency service finder for İzmir with 4,600+ verified entries. "Open Now" filtering checks current time against business hours. GPS-based distance sorting, mobile-friendly map, and quick-action buttons for WhatsApp, calling, and Google Maps directions.
Marketing site for the Confluence Array indicator. Galaxy canvas animation with twinkling stars and nebula glows, animated typing hero, Pleiades star cluster SVG logo (hand-coded), pricing toggle, live preview integration, and full SEO setup with custom domain on GitHub Pages.
ML regression pipeline on the Ames Housing Dataset (80 variables, 3k properties). Compared Linear Regression, Ridge, Random Forest, and XGBoost after feature engineering (TotalSF, TotalBaths, HouseAge) and log target transformation. XGBoost achieved R² 0.944.
Analyzed 540k transactions from a UK wholesale gift retailer. EDA, Pareto analysis (top 26% of customers → 80% of revenue), RFM segmentation into Champions / At-Risk / Lost cohorts, and a cohort retention heatmap revealing 20.6% month-1 retention and 33% churn.