Cagan Oflazoglu · econometrics and machine learning · Amsterdam

Now: Pre-Master's programme in Econometrics at the University of Amsterdam, and a researcher on the Commodities Desk at Prometheus Capital, a student-run investment fund in Amsterdam.

Recent work: a free trainer for the Optiver online assessment, a camera-roll cleaner whose photos never leave your own devices, and an open Turkish speech model in development.

• OptPrep (repo): free practice for the Optiver online assessment. Seven tasks on their real clocks and nine Zap-N games; 3,094 generated questions re-checked by a C++20 engine (order-book branch and bound, Monte Carlo) with zero failures; 142 lessons, and every wrong option names the misconception that leads to it. 793 tests, no runtime dependencies, works offline. npx optprep. JavaScript, C++, Python.
• Phone Memory Slider: cleanup for your camera roll that stays on your own devices. A preference model trained on your favourites ranks duplicates, screenshots and blurry shots, you swipe to decide, and anything you drop goes to Android's recoverable trash. Pairing runs over TLS with a pinned certificate; no account, no cloud. Beta. C++23, Qt, Kotlin.
• Liminal: meeting minutes in Romanian, Russian and English for Medpark, Moldova's largest private hospital, from meetings that switch language mid-sentence, on a server with no connection out. 2nd place, DeepTech GigaHack 2026 HealthTech track. 100% of decisions found and correct, with no trap fallen for across 19 model runs; 93% speaker accuracy, live, on a laptop CPU; character error cut to 20.6% by decoding every utterance twice; 506 tests. Team of five; I worked on the AI/ML and backend. Python.
• ML Factory: give it a CSV and a target column. An agent runs 33 statistical checks, ranks about 3,000 pipelines, races a 27-config shortlist in parallel on Modal, and confirms the winner on 10 paired folds with corrected t and Bayesian tests against a holdout locked before any diagnostic runs. Built for the Accel Hackathon. Python.
• CamToParkingSlot: parking search that matches RDW vehicle dimensions against Amsterdam's 210,247 surveyed bays and reads 69 live street cameras for free spaces: 0.00% false fits, 170 ms search. Python, C++.
• VC-Alpha-Signal-Finder: evaluation engine for early-stage venture. Reads seven public sources, builds founder and company dossiers, and scores each against a plain-English thesis with the evidence attached; a browser extension sends the page you are reading to the engine and shows its verdict. Python.

turkish-sa-asr (private until release): Turkish speech recognition that also says who spoke. The bar is the best published Turkish score, 4.71% WER on FLEURS-tr from a closed API; the best open model scores 5.77% in the project's own harness. So far: 966.5 hours of training speech checked against every test set, 1,519.8 hours of parliament audio aligned to the official record, six open models scored on four sealed test sets.

Lou (controlled legal memory: contract clauses mapped to a playbook, 63 tests; built for the START Munich hackathon, Siemens track) · Was-Liberation-a-Delusion (BSc thesis: PPML gravity model on the 2025 US tariffs, with the full replication pipeline) · Octant-AI (agent that turns an investment thesis into a financial model and a research report) · calibrcv-cli (offline resume optimizer with LaTeX output) · Portfolio-Manager (trades only on a unanimous five-model vote plus a risk gate; work in progress) · TCDD-Koltuk-Bul (catches freed-up seats on sold-out Turkish State Railways trains) · mental-math-forge (mental arithmetic trainer) · neilization (Claude skill for public-science prose)

LinkedIn