Senior AI researcher and AI architect. I build small AI models and tools for decisions whose confidence and evidence people can inspect. I founded Eschatia Labs, an international open-source research lab, and work as an AI Engineer at Enocta.
Visit my site for project demos and research. Hugging Face hosts the models and datasets.
- JevAlt: small language models for choosing between options in English, Turkish and German, with a probability for each answer. Karar-4B is one of the best open Turkish decision models at 4B parameters: 96.8% accuracy on held-out Turkish decisions against Kev-4B's 87.1%. The tests come from JevAlt's own data pipeline. Try the models.
- JevOss: test a decision model's accuracy, confidence and response to hidden instructions. Test methods.
- Open benchmarks: jevalt-bench has 3,116 written situations in English, Turkish and German for decision models, and Tholos-Bench has 160 workspace tasks for small AI assistants. Both are free to download and run against your own model.
- xdfdet: eight deepfake video detectors with heatmaps showing the face regions behind a prediction, from my MSc thesis. Study and demos.
- Ultra Mod: an all-in-one mod pack for Claude Code. Usage limits and context above the prompt, a guard with undo for
rm -rfandgit reset --hard,.envfiles kept away from the model, and a receipt under every answer. Install withnpx ultramod. Site and demos. - Tholos: AI assistants that share tables, notes and tasks on your computer, with rules that allow, ask about or deny actions. Tholos-2B passed 137 of 160 Tholos-Bench scenarios against MiniCPM5-2B's 112 on a Kaggle T4 with llama.cpp and JSON schema decoding. Model and benchmark.
- reevesagents: run AI coding tools side by side in a terminal workspace, or let one direct the others. Demo and docs.
MSc Computer Science, TED University. BSc Computer Engineering, TOBB University of Economics and Technology.
- M. Kaya, V. Adanova. Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention. UBMK 2026. Accepted, to appear in IEEE Xplore.
- M. Kaya. Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques. MSc thesis, TED University, 2025. Advisor: V. Adanova. Read the thesis.
- M. Kaya. Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed. Preprint, Eschatia Labs, 2026. Read the paper.