Source: https://eschatialabs.com/research/ # Research Everything we publish, grouped by area. [Deepfake detection](https://eschatialabs.com/research/#deepfake-detection) [Decision models](https://eschatialabs.com/research/#decision-models) [Local AI and open tools](https://eschatialabs.com/research/#local-ai) [Language preservation](https://eschatialabs.com/research/#language-preservation) [Public benchmarks](https://eschatialabs.com/research/#benchmarks) ## Deepfake detection Detectors that show where they look and how far their scores can be trusted. ### Publications - [Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention](https://eschatialabs.com/research/ubmk-2026/) Mert Kaya, Venera Adanova. An EfficientNet-B4 deepfake detector trained under nine augmentation and cutout configurations on FaceForensics++, scored on AUC, F1, Brier score and log loss, with Grad-CAM attention over facial regions. The four metrics do not agree on one best configuration. Accepted at UBMK 2026. The final version will appear in IEEE Xplore. - [Read](https://eschatialabs.com/research/ubmk-2026/) - [PDF](https://eschatialabs.com/papers/ubmk-2026.pdf) - [Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques](https://eschatialabs.com/research/msc-thesis/) Mert Kaya. A master’s thesis comparing augmentation and cutout techniques for frame-level CNN deepfake detection, with Grad-CAM analysis of attention over facial regions. MSc thesis, TED University, 2025. - [Read](https://eschatialabs.com/research/msc-thesis/) - [PDF](https://eschatialabs.com/papers/msc-thesis.pdf) - [DOI](https://doi.org/10.5281/zenodo.18998566) ### Projects - [xdfdet](https://xdfdet.mertkayacs.com/) Code and eight trained EfficientNet-B4 deepfake detectors from the thesis, with Grad-CAM region analysis and a public site with interactive detection and estimation games. Code: MIT. Weights: CC BY-NC 4.0, under FaceForensics++ terms. - [Real or AI?](https://xdfdet.mertkayacs.com/game/) - [GitHub](https://github.com/mertkayacs/xdfdet) - [Hugging Face](https://huggingface.co/mertkayacs/xdfdet) - [Kaggle](https://www.kaggle.com/models/mertilovski/xdfdet) ## Decision models Small models that answer questions in English, Turkish and German with calibrated probabilities, on a laptop CPU. ### Publications - [JevAlt: Technical report](https://eschatialabs.com/research/jevalt-report/) Mert Kaya. Method, held-out results and limits of the open English, Turkish and German decision models. Technical report (not peer-reviewed) - [Read](https://eschatialabs.com/research/jevalt-report/) - [PDF](https://eschatialabs.com/papers/jevalt-report.pdf) ### Projects - [JevAlt](https://jevalt.mertkayacs.com/) Three open decision models for English, Turkish and German. Each answer comes with a calibrated probability, and the models run in about 3 GB of RAM. - [GitHub](https://github.com/mertkayacs/jevalt) - [Hugging Face](https://huggingface.co/collections/mertkayacs/jevalt-6abde16559675245349acea0) - [Demo](https://huggingface.co/spaces/mertkayacs/JevAlt) - [Kaggle](https://www.kaggle.com/models/mertilovski/jevalt) - [Emberwick](https://emberwick.mertkayacs.com/) A medieval village in the browser where the JevAlt models decide how villagers respond to storms, fires, markets and wolves at dusk. ### Models - [Deem-4B](https://huggingface.co/mertkayacs/Deem-4B) English, with GGUF builds. - [GGUF](https://huggingface.co/mertkayacs/Deem-4B-GGUF) - [Karar-4B](https://huggingface.co/mertkayacs/Karar-4B) Turkish, with GGUF builds. - [GGUF](https://huggingface.co/mertkayacs/Karar-4B-GGUF) - [Wähler-4B](https://huggingface.co/mertkayacs/Wahler-4B) German, with GGUF builds. - [GGUF](https://huggingface.co/mertkayacs/Wahler-4B-GGUF) ### Datasets - [jevalt-data](https://huggingface.co/datasets/mertkayacs/jevalt-data) Training data for the JevAlt decision models. - [emberwick-videos](https://huggingface.co/datasets/mertkayacs/emberwick-videos) Recorded scenes and films from Emberwick. ## Local AI and open tools Agents and tools that run on your own machine. ### Publications - [Tholos-2B: Technical report](https://eschatialabs.com/research/tholos-2b-report/) Mert Kaya. Training method, Tholos-Bench results and limits of the small local agent model. Technical report (not peer-reviewed) - [Read](https://eschatialabs.com/research/tholos-2b-report/) - [PDF](https://eschatialabs.com/papers/tholos-2b-report.pdf) ### Projects - [Tholos](https://tholos.mertkayacs.com/) Small agents sharing a local workspace of tables, notes and a task board, with allow, ask and deny rules for each action. - [GitHub](https://github.com/mertkayacs/tholos) - [reevesagents](https://reevesagents.mertkayacs.com/) Open source software development. A local tmux workspace for coding agents, with a CLI, terminal interface, web interface and MCP server in one Apache-2.0 npm package. - [GitHub](https://github.com/mertkayacs/reevesagents) - [npm](https://www.npmjs.com/package/reevesagents) ### Models - [Tholos-2B](https://huggingface.co/mertkayacs/Tholos-2B) A 2B agent model fine-tuned from MiniCPM5-2B on executed trajectories, released under Apache-2.0 with GGUF builds. - [GGUF](https://huggingface.co/mertkayacs/Tholos-2B-GGUF) ### Datasets - [tholos-trajectories](https://huggingface.co/datasets/mertkayacs/tholos-trajectories) Executed agent trajectories used to train Tholos-2B. ## Language preservation One of the best research efforts on language preservation with small local models. ### Projects - [Eldalambë](https://eldalambe.mertkayacs.com/) A small language model that reads and writes Quenya and Sindarin, including Tengwar script, grounded in a grammar-verified corpus, with tools for translation, lookup, inflection and Tengwar rendering. Code and weights are private. Demo by invitation. ## Public benchmarks Open tests and their results. - [JevOss](https://jevoss.mertkayacs.com/) An open test bench for decision models using the Jev API, measuring accuracy, probability calibration and failure cases. - [GitHub](https://github.com/mertkayacs/jevoss) - [jevalt-bench](https://huggingface.co/datasets/mertkayacs/jevalt-bench) Benchmark and live-test results for JevAlt and the comparison models. - [Tholos-Bench](https://github.com/mertkayacs/tholos#tholos-bench) 160 scenarios for local agents, each scored on the final state of the workspace.