AI agents · Data engineering · Operations research
I build Python systems that connect language models, data workflows and mathematical optimization. At Cravio, I work on founding and technical development, with experience in FastAPI, RAG and agent workflows. My academic background includes an MSc in Industrial Engineering at Özyeğin University.
My toolkit includes Python, FastAPI, Docker, Airflow, PostgreSQL and Gurobi. I am interested in opportunities where reliable software and explicit decision models meet.
USA Energy Market is a knowledge base and teaching system for US electricity and gas markets, built with AI coding agents under strict, test-enforced rules. It turns energy podcasts, papers and grid-operator documents into cited evidence and linked notes, then into bilingual lessons and calculators.
- Pipeline: RSS feeds, publisher transcripts and local speech-to-text (Parakeet on MLX) for 1,000+ episodes, plus open-access papers and reports.
- Quote gate: every mined claim, number and prediction must quote its source exactly, or it fails validation.
- Teaching: 71 lessons in English and Turkish, 13 calculators (LMP, capacity auctions, FTRs, spark spreads and more), and an Obsidian vault of 6,000+ notes.
- Market data:
usgrid, an eptr2-style client overgridstatusfor the seven US grid operators.
What's next (full roadmap):
| When | Plan |
|---|---|
| Nov–Dec 2026 | PJM and FERC EQR data through PUDL, daily data snapshots, primary-source lesson audits, weekly podcast sync |
| Q1 2027 | Settlement rule tables, usgrid on PyPI, a battery arbitrage backtester, the dashboard on GitHub Pages |
| Q2 2027 | FTR and congestion analytics, a capacity auction model, live market panels in lessons |
| H2 2027 | Interconnection-queue and data-center tracker, forecast baselines, a podcast prediction scoreboard, a cited Q&A agent and MCP server |
Three free, open-source deck builders that share one design: a pure TypeScript engine, SQLite, card images on disk and a Next.js interface on localhost. No account and no AI at runtime; every score is computed from card text, rules and synced data, and is shown with its reasons. Each ships as a portable download for macOS and Windows.
| Tool | Game | What it does | Download |
|---|---|---|---|
| Commander Forge | Magic: The Gathering (Commander) | Bracket-aware auto-build, 0–100 card scores per commander, deck rating, improve suggestions, printable proxies, a deck finder for your collection. |  |
| Riftbound Forge | Riftbound | Auto-builds legal decks, scores every card, rates a list, finds synergies and win conditions, gives matchup and side-deck lines for 1v1, 2v2 and FFA. |  |
| Snap Forge | Marvel Snap | Brews off-meta decks that still clear a power floor, with a deck rating, a separate novelty score against synced meta lists, energy curve and draw odds, and swap suggestions. |  |
These are unofficial fan projects built with AI coding agents in staged, test-gated workflows. Ratings are heuristic construction checks, not win-rate predictions; each repository states what is measured and what is not. Card data and images are downloaded to the user's machine and are not part of the repositories.
| Project | What to explore |
|---|---|
| Agent Lot Sizing | SmolAgents and a local language model connected to deterministic Gurobi tools; separate capacitated and uncapacitated planning models. |
| Multi-Item Perishable Lot Sizing | Compact mixed-integer formulations and branch-and-price experiments for production capacity and shelf-life constraints. Research code with several model iterations. |
| Coil Stacking with Q-Learning | A custom stacking environment and tabular Q-learning experiment with support, blocking and urgency rules. An experimental policy, with validation work still ahead. |
| Explainable Road-Repair Prioritization | An Ataşehir prototype combining OSM roads, simulated damage/POIs, XGBoost and SHAP. Heuristic training targets and distance-unit validation remain research limitations. |
- EPİAŞ Price Explorer — a Streamlit date selector and table viewer built around the external
eptr2electricity-market data client. - Mechabellum Build Assistant — a fan tool for counter lookups, composition vulnerabilities and tier-aware suggestions using structured game data.
- Galio's Den — a Java console dungeon game with randomized encounters, resource choices and boss fights.
I use side projects to turn an idea into an interface, work with imperfect data and explore game mechanics. Repository maturity varies; research prototypes should be read with their assumptions and limitations.
eptr2 and Catan-AI are forks of upstream projects. They are separate from the original projects above; these links do not imply authorship of the upstream code.
You can explore my work and contact me through GitHub.