Mert Kaya
MenuClose

Mert Kaya AI research and engineering on small open models, explainable deepfake detection and local AI tools

I build small AI models for specific jobs, open tests that measure them and tools that show why a model made its decision.

Founder of Eschatia LabsAI Engineer at Enocta

Selected workPublications

8,600model and dataset downloads on Hugging Face

9,000downloads on npm

150GitHub stars

As of

Selected work

All work

JevAlt

Small models that read a situation, pick one of the given answers and give a probability for each answer. Deem-4B works in English, Karar-4B in Turkish and Wähler-4B in German.

All three were trained from Intern-Decision-4B, first on the three languages together, then on each language alone.

On test questions kept apart from training, Deem-4B reaches 94.7% accuracy, versus 84.7% for Kev-4B. Karar-4B reaches 96.8%, versus 87.1% for Kev-4B, and Wähler-4B reaches 92.0%, versus 81.1% for Kev-4B.

Related:JevOssEmberwick

Tholos + Tholos-2B

Tholos lets AI assistants carry out tasks on your own computer, in a shared workspace with tables, notes and a task board. For each action, rules decide whether it runs, waits for your approval or is blocked.

Tholos-2B is the small model behind the assistants, trained from MiniCPM5-2B on recordings of completed tasks.

On the same graphics card (a Kaggle T4), Tholos-2B passes 137 of the 160 Tholos-Bench tasks, versus 112 for MiniCPM5-2B. On the same ordinary processor (CPU), Tholos-2B passes 134, versus 117 for MiniCPM5-2B.

Apache-2.0 license.

reevesagents

Runs several AI coding assistants side by side in one terminal, with one assistant coordinating the others.

One npm package, Apache-2.0 license, 10 interface languages.

Ultra Mod

The all-in-one mod pack for Claude Code, Anthropic’s coding assistant. It shows your usage limits and how full the context window (the model’s working memory) is, asks before rm -rf or git reset --hard runs and can undo either, keeps .env files and other secrets away from the model, and adds a receipt under every answer.

MIT license. No network calls. About 0 always-on model tokens.

xdfdet

The eight trained detectors from the MSc thesis, with code, heat maps and two browser games. The related paper is accepted at UBMK 2026.

Eight released models. Code: MIT; trained model files: CC BY-NC 4.0 under the terms of FaceForensics++, a public collection of real and altered face videos.

Related:UBMK 2026 paperMSc thesis

More projects: Eldalambë. Also on the Work page: datasets and early work. The open benchmarks have their own pages.

Open source

Full open source list

A short list of contributions to external projects.

  • musistudio avatar
    musistudio/claude-code-router 37,723
    PR
  • sgl-project avatar
    sgl-project/sglang 37,262
    PR
  • jarrodwatts avatar
    jarrodwatts/claude-hud 28,618
    PR
  • EleutherAI avatar
    EleutherAI/lm-evaluation-harness 14,253
    PR

Publications

All research

Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention

Mert Kaya, Venera Adanova

UBMK 2026, 11th International Conference on Computer Science and Engineering, Istanbul (IEEE)

Accepted, to appear in IEEE Xplore.

An accepted paper comparing nine ways to train a deepfake detector: how accurate it is, how well its confidence matches its results, and which parts of the face it relies on.

Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques

Mert Kaya, Advisor: Venera Adanova

MSc thesis, TED University,

The thesis compares ways to train models that detect AI-altered faces in video. It uses heat maps to show which facial regions affected each result.

Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed

Mert Kaya

Public preprint, not peer-reviewed. Eldalambë's code, data, models and benchmark are private.

This paper studies Eldalambë, an assistant for Tolkien’s invented Elvish languages, Quenya and Sindarin. The adopted version, Pro-Preserve, writes the expected sentence for 40 of 100 test requests, versus 21 for the earlier version, Pro2508, and identifies Tolkien’s words in 277 of 300 cases, versus 259 for Pro2508.

The work in numbers

Every claim on this site links to the benchmark, the data or the code behind it.

Karar-4B, Turkish
96.8% correct on test questions excluded from training; the comparison model Kev-4B reaches 87.1%. Source: jevalt-bench.
Deem-4B and Wähler-4B, English and German
94.7% and 92.0% correct on their own languages, against 84.7% and 81.1% for Kev-4B. Source: jevalt-bench.
Tholos-2B, assistant tasks
On the same graphics card (a Kaggle T4), Tholos-2B passes 137 of 160 Tholos-Bench tasks, versus 112 for MiniCPM5-2B. On the same ordinary processor (CPU), it passes 134, versus 117 for MiniCPM5-2B. Source: Tholos-Bench.
reevesagents, agent workspace
107 GitHub stars as of October 15, 2026 and 742 npm downloads from September 13 to October 12, 2026. Source: GitHub and npm.
Citation (BibTeX)
@mastersthesis{kaya2025explainable,
  author = {Kaya, Mert},
  title = {Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques},
  school = {TED University},
  year = {2025},
  doi = {10.5281/zenodo.18998566}
}

Last updated:

Frequently asked questions

Who is Mert Kaya?

Mert Kaya is a senior AI researcher and AI engineer, and the founder of Eschatia Labs, an independent open-source AI research lab. He builds small open models and local AI tools whose claims can be tested: on the jevalt-bench Turkish test questions excluded from training, his Karar-4B model answers 96.8% correctly, against 87.1% for the comparison model Kev-4B; his open decision-model family covers English, Turkish and German.

What is reevesagents?

reevesagents is Mert Kaya’s open-source tool for running AI coding agents side by side: Claude Code, Codex, Kimi, OpenCode and other assistants work in one terminal workspace, with one agent coordinating the team. The Apache-2.0 tool held 107 GitHub stars on October 15, 2026 and logged 742 npm downloads from September 13 to October 12, 2026.

What is xdfdet?

xdfdet is Mert Kaya’s system for detecting AI-altered faces in video, built from his MSc thesis at TED University. It publishes the full code and eight trained image models, with heat maps that show which facial regions affected each result. The code uses the MIT license, and the thesis is public on Zenodo.

What is Karar-4B?

Karar-4B is Mert Kaya’s open decision model for Turkish. On the jevalt-bench test questions excluded from training it answers 96.8% correctly, against 87.1% for the comparison model Kev-4B, and it runs locally in about 3 GB of memory. On the same jevalt-bench test questions, its German and English siblings Wähler-4B and Deem-4B reach 92.0% and 94.7% on their own languages, against 81.1% and 84.7% for Kev-4B.

What is Tholos?

Tholos is Mert Kaya’s workspace for AI assistants that run on your own computer. They share tables, notes and a task board. For each action, rules decide whether it runs, waits for your approval or is blocked. On the same graphics card (a Kaggle T4), its task model Tholos-2B passes 137 of 160 Tholos-Bench tasks, versus 112 for MiniCPM5-2B, the model it was trained from. On the same ordinary processor (CPU), Tholos-2B passes 134, versus 117 for MiniCPM5-2B.

Where can I find and cite Mert Kaya’s research?

Mert Kaya’s MSc thesis on explainable deepfake detection is public on Zenodo with the DOI 10.5281/zenodo.18998566, and his comparative study of augmentation and cutout methods was accepted at UBMK 2026 and will appear in IEEE Xplore. Every publication page on this site carries a BibTeX record you can copy, and the code and models sit on GitHub and Hugging Face.

Eschatia Labs

Founded by Mert Kaya

Eschatia Labs is an independent open-source AI research lab.

eschatialabs.com

Experience

Enocta, the Credit Bureau of Turkey (KKB), AgriDL5.0 and Alictus.

Experience and education

Contact

Email me about research or engineering work.

[email protected]