I like problems where the data is messy and the answer has to be right: tomorrow's electricity price, a delivery van in Istanbul traffic, a collapsed building in a satellite image, a bank statement full of shop names nobody has labelled.
My loop is usually the same. I find a public data source that nobody has cleaned up yet (EPİAŞ, İBB, KAP, Eurostat, OpenF1), build a model, and test it on data it has never seen against the best baseline that already exists. Then I ship it as something you can open in a browser, with a daily pipeline that keeps it current. Every README says where the model falls short, because a number you cannot check is not worth much.
I graduated first in my faculty at Istanbul Ticaret University, and I have shipped a market analysis app to the App Store.
| Models forecasting · probabilistic & calibrated outputs · Monte Carlo · naive Bayes · CNNs · RAG |
Data Turkish open data · geospatial · time series · PDF and table extraction · daily pipelines |
Product dependency-free web apps · SVG charts · PWAs · Flutter · FastAPI · CI and tests |
PERSONAL FINANCE · ON-DEVICE ML
Bank statement analysis that never leaves the browser. A naive Bayes model trained in the browser categorises local shops that brand rules miss, and asks when it is unsure. 14 % → 79 % correct on 121 unseen merchants · 97 % precision when it commits · personal inflation · savings planner  
|
ENERGY MARKETS · FORECASTING
Day-ahead probabilistic forecasts of Turkish electricity prices, demand, wind, solar and carbon, with Shapley explanations and an auction simulator on EPİAŞ's real bids. Beats TEİAŞ's load plan by 25 % and EPİAŞ's wind forecast by 34 % · a battery on it earns 82 % of perfect foresight  
|
LOGISTICS · OPTIMISATION
Tomorrow's street-level traffic forecast from İBB open data, fed into a time-dependent ALNS optimiser written from scratch, then driven on the traffic that really happened. 10–17 % fewer driver hours than Google OR-Tools at the same punctuality · traffic error 3.5 km/h vs 9.3  
|
SPORTS ANALYTICS · SIMULATION
Every 2026 Formula 1 race rebuilt from raw timing data, with a live win probability, a strategy simulator and a tyre model. It updates itself after every race. Tyre model 28 % lower error than a flat-pace baseline on 9 unseen races · 380 / 380 championship points reconciled  
|
COMPUTER VISION · DISASTER RESPONSE
Building-level damage from pre- and post-event satellite images of the 6 February 2023 earthquakes, tested on Antakya, a city kept out of training entirely. The top 10 % of its ranking finds 51–55 % of destroyed buildings (10 % for a change score) · calibrated probabilities  
|
LLM AGENTS · RETRIEVAL
Answers questions about Borsa İstanbul companies from their KAP filings, with page-level citations. Ratios come from extracted statement values, never from model arithmetic. Retrieval Hit@5 0.912 (BM25 alone 0.692) · Turkish questions 0.961 · 90 / 90 statement values exact 
|
 |
iOS · ANDROID · ON THE APP STORE
Market analysis and paper trading. A Flutter client and a FastAPI backend with real-time WebSocket market data, portfolio risk and diversification analytics, ML and LLM market insights, and PDF reports. Educational simulation, no real money.    
|
Software Testing Intern · EGEMSOFT / Türk Telekom
Performance testing with JMeter, Oracle database testing, found critical deployment issues, built a data mining demo in Python.
Software Intern · Kazancı Holding
Multi-layer .NET architecture, MS SQL Server integration, Docker deployment.
Istanbul Ticaret University
Ranked 1st in the faculty, GPA 3.54 / 4.00.
Earlier projects
- Vehicle image classification: transfer learning with EfficientNet
- Fish species classifier: a CNN tuned with Keras Tuner
- Android chat app: real-time messaging with Firebase authentication (Java)
- Client-server chat: secure messaging on Windows (C#, .NET)
- Cargo tracking system and a logistics website
- iOS apps: machine learning on iOS, brain training