I build things that have to be right: production software and quant research that survives its own tests.
Final-year Computer Science at the University of Sheffield. I ship a real product (Tayf, a Turkish news bias analyser running on Supabase + Vercel) and write portfolio-theory and trading research from first principles, with the out-of-sample numbers reported honestly.
Tayf
AI/ML news-analysis platform for a free and unfiltered press: same news, different worlds. A real-time pipeline over 144 Turkish outlets runs a three-method ML story-clustering ensemble (TF-IDF similarity, 4-gram fingerprints, entity overlap), news sentiment analysis, who-said-what views across bias zones, plus blindspot and cross-spectrum surprise detection. ~2,200 monthly site visits as of June 2026. Early stage, stated plainly.
markowitz-optimizer
From-scratch mean-variance optimization reproducing canonical literature: efficient frontier, Black-Litterman, Ledoit-Wolf.
pairs-trading
Cointegration stat-arb lab. The out-of-sample number is the headline, judged by Deflated Sharpe, not in-sample.
ma-crossover-backtest
Vectorised MA-crossover backtester on US-equity ETFs. The point is rigorous evaluation, not a winning strategy.
Identifying Asset-Price Exuberance and Negative Abnormal Returns in UK Equities via Machine Learning
Can exuberance, momentum and volatility features flag UK equities likely to deliver negative abnormal returns out of sample? Mostly no. That is the finding. Rejections collapsed from 14/74 to 2/74 across variance estimators, and the long-short signal failed the deflated-Sharpe test. Reported as an audited predictability claim, not a trading strategy.
- ▸Leakage-safe walk-forward pipeline; hyperparameters committed to a SHA-256 manifest before evaluation
- ▸Dependence-aware multiple testing (Clark-West, Romano-Wolf, Model Confidence Set) over an iid/HAC/Politis-White variance envelope
- ▸O(T²) prefix-sum BSADF for exuberance detection
- ▸Long-short signal failed the deflated-Sharpe test; reported anyway
Software Hut Prize (Client)
Team judged by the client to have delivered the most effective software under an agile process. Group lead for an AI-assisted Kanban board built for the consultancy Reply (COM3420).
BetterPicked · Engineering "You're Hired"
An AI / computer-vision system to cut supermarket food waste by classifying overripe and misshapen produce: ~40% discard reduction via a CNN. Led financial modelling, market research, and the investor pitch.
Final-year Computer Science student at the University of Sheffield (BSc Hons, 2023-2026) and co-founder of Tayf, an AI/ML news-analysis platform running a real-time pipeline over 144 Turkish outlets. My dissertation asks whether exuberance plus momentum and volatility features can flag UK equities headed for negative abnormal returns out of sample. The honest answer: mostly they can't. Rejections collapsed from 14/74 to 2/74 under dependence-aware testing, and the long-short signal failed the deflated-Sharpe test. The same standard runs through twenty ML-finance projects rebuilt from first principles, ~1,200 tests in total. Every claim is judged out of sample. Negative results ship with the same confidence as positive ones. Bilingual in English and Turkish.