# Work

AI models, software and datasets for choosing among fixed answers, helping assistants complete tasks and detecting AI-altered faces. Each entry says what is public.

- [Models](https://mertkayacs.com/work/#models)

- [Publications](https://mertkayacs.com/work/#papers)

- [Open-source software](https://mertkayacs.com/work/#software)

- [Simulations](https://mertkayacs.com/work/#simulations)

- [Open benchmarks](https://mertkayacs.com/work/#benchmarks)

- [Datasets](https://mertkayacs.com/work/#datasets)

- [Early work](https://mertkayacs.com/work/#early)

- [Infrastructure](https://mertkayacs.com/work/#infrastructure)

## Models

### [JevAlt](https://jevalt.mertkayacs.com)

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.

Runs on a laptop’s main processor (CPU) in about 3 GB of memory per model. Compare results with Kev-4B below.

- [Visit project](https://jevalt.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/jevalt)

- [Model collection](https://huggingface.co/collections/mertkayacs/jevalt-decision-models)

- [Try the demo](https://huggingface.co/spaces/mertkayacs/JevAlt)

- [Kaggle](https://www.kaggle.com/models/mertilovski/jevalt)

### [Deem-4B](https://huggingface.co/mertkayacs/Deem-4B)

Deem-4B answers fixed-option questions in English and estimates each answer’s probability.

Correct answers on the jevalt-bench English questions excluded from training: Deem-4B 94.7%, Kev-4B 84.7%.

- [Hugging Face](https://huggingface.co/mertkayacs/Deem-4B)

- [Run on your computer (GGUF)](https://huggingface.co/mertkayacs/Deem-4B-GGUF)

### [Karar-4B](https://huggingface.co/mertkayacs/Karar-4B)

Karar-4B answers fixed-option questions in Turkish and estimates each answer’s probability.

Correct answers on the jevalt-bench Turkish questions excluded from training: Karar-4B 96.8%, Kev-4B 87.1%.

- [Hugging Face](https://huggingface.co/mertkayacs/Karar-4B)

- [Run on your computer (GGUF)](https://huggingface.co/mertkayacs/Karar-4B-GGUF)

### [Wähler-4B](https://huggingface.co/mertkayacs/Wahler-4B)

Wähler-4B answers fixed-option questions in German and estimates each answer’s probability.

Correct answers on the jevalt-bench German questions excluded from training: Wähler-4B 92.0%, Kev-4B 81.1%.

- [Hugging Face](https://huggingface.co/mertkayacs/Wahler-4B)

- [Run on your computer (GGUF)](https://huggingface.co/mertkayacs/Wahler-4B-GGUF)

### [xdfdet](https://xdfdet.mertkayacs.com)

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.

- [Visit project](https://xdfdet.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/xdfdet)

- [Hugging Face](https://huggingface.co/mertkayacs/xdfdet)

- [Kaggle](https://www.kaggle.com/models/mertilovski/xdfdet)

### [Eldalambë](https://eldalambe.mertkayacs.com)

Eldalambë helps people read and write Tolkien’s invented Elvish languages, Quenya and Sindarin, and their script, Tengwar. A small language model checks its answers against a dictionary and grammar rules.

Demo by invitation. Code, data, trained model files and benchmark are private; only the paper is public.

- [Visit project](https://eldalambe.mertkayacs.com)

- [Paper (preprint)](https://eschatialabs.com/research/eldalambe-2026/)

### [Tholos-2B](https://huggingface.co/mertkayacs/Tholos-2B)

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

Apache-2.0 license.

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.

- [Hugging Face](https://huggingface.co/mertkayacs/Tholos-2B)

- [Run on your computer (GGUF)](https://huggingface.co/mertkayacs/Tholos-2B-GGUF)

## Publications

### [Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention](https://mertkayacs.com/research/#ubmk-2026)

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.

Accepted at UBMK 2026. The final version will appear in IEEE Xplore; no DOI has been assigned.

### [Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques](https://mertkayacs.com/research/#msc-thesis)

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.

MSc thesis, TED University, 2025. Advised by Venera Adanova.

- [Thesis on Zenodo](https://doi.org/10.5281/zenodo.18998566)

### [Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed](https://mertkayacs.com/research/#eldalambe-2026)

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.

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

- [Read the paper](https://eschatialabs.com/research/eldalambe-2026/)

- [PDF](https://eschatialabs.com/papers/eldalambe-2026.pdf)

## Open-source software

### [reevesagents](https://reevesagents.mertkayacs.com)

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.

 7,400 downloads on npm

   As of October 18, 2026

- [Visit project](https://reevesagents.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/reevesagents)

- [npm](https://www.npmjs.com/package/reevesagents)

### [Ultra Mod](https://ultramod.mertkayacs.com)

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.

- [Visit project](https://ultramod.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/ultramod)

- [npm](https://www.npmjs.com/package/ultramod)

### [Tholos](https://tholos.mertkayacs.com)

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.

- [Visit project](https://tholos.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/tholos)

## Simulations

### [Emberwick](https://emberwick.mertkayacs.com)

A medieval village in your browser where the three decision models choose how villagers react to storms, fires, markets and wolves.

- [Visit project](https://emberwick.mertkayacs.com)

- [Videos](https://huggingface.co/datasets/mertkayacs/emberwick-videos)

## Open benchmarks

### [JevOss](https://eschatialabs.com/benchmarks/jevoss/)

Test whether a decision model answers correctly and gives reliable probabilities, using a common request format (Jev API).

- [Visit project](https://jevoss.mertkayacs.com)

- [GitHub](https://github.com/mertkayacs/jevoss)

### [jevalt-bench](https://eschatialabs.com/benchmarks/jevalt-bench/)

Test questions and results comparing JevAlt’s decision models with other models.

- [Hugging Face](https://huggingface.co/datasets/mertkayacs/jevalt-bench)

### [Tholos-Bench](https://eschatialabs.com/benchmarks/tholos-bench/)

160 tests of AI assistants, scored by whether the completed work matches the task.

- [GitHub](https://github.com/mertkayacs/tholos/tree/main/tholos/bench)

- [Hugging Face](https://huggingface.co/datasets/mertkayacs/tholos-bench)

## Datasets

### [jevalt-data](https://huggingface.co/datasets/mertkayacs/jevalt-data)

Training examples for JevAlt’s English, Turkish and German decision models.

- [Hugging Face](https://huggingface.co/datasets/mertkayacs/jevalt-data)

### [emberwick-videos](https://huggingface.co/datasets/mertkayacs/emberwick-videos)

Recorded scenes from Emberwick, the browser village simulation.

- [Hugging Face](https://huggingface.co/datasets/mertkayacs/emberwick-videos)

### [tholos-trajectories](https://huggingface.co/datasets/mertkayacs/tholos-trajectories)

Recorded task executions used to train Tholos-2B, the model for Tholos’s AI assistants.

- [Hugging Face](https://huggingface.co/datasets/mertkayacs/tholos-trajectories)

## Early work

### Turkish poem generator

A deep-learning model trained on a collection of Turkish poems for a university course on computer processing of language.

Graded A. Code and paper are private.

### AgriDL smart agriculture

A mobile app applying deep learning to farming, using sensor readings. Built under contract with government partners.

Code is private.

## Infrastructure

### Self-hosted production platform

Runs the public projects, monitors services, sends alerts and keeps encrypted off-site backups. Network protections and setup are defined in code.

Private.

Canonical: https://mertkayacs.com/work/