I build models that know how sure they are. Accuracy tells you how often a model is right. Calibration tells you when you can trust it. My research sits at that boundary, and the open-source Kairos models are where it becomes working code.

I am an ML/NLP/LLM researcher and engineer. For almost two years I have worked full time on machine learning, and in that time I moved from classical models to the frontier of generative modeling and language model research. I design models, train them, publish them and write about them.

My work is built around one principle: a model must be accurate, and it must also know how confident it is. That principle is the core of my research direction. It led me to CIDA, to a series of papers on calibration and multi-agent systems, and to the open-source Kairos models.

I combine research depth with engineering discipline. I understand how modern architectures work from the inside, and I can take an idea all the way to a trained, evaluated and published model. Alongside research I compete in algorithmic programming and lead my university ICPC team, which keeps my implementation skills sharp and my thinking about complexity honest.

I am developing the next generation of the Kairos model family. It combines CIDA with the System One decision-model approach that Jev and Laya brought into the field: compact models that return structured decisions with calibrated probabilities instead of free-form text. Such models are fast, cheap to run, and suited to agent pipelines where every decision must be measurable and trustworthy.

I am writing papers on CIDA, calibrated prediction and multi-agent systems, and I present this work at conferences. Each paper feeds the next Kairos release, and each release gives the papers something concrete to measure.

flowchart LR
    A["Qwen1.5 base models<br/>1B and 7B"] --> B{{"CIDA method"}}
    B --> C["Kairos 1B<br/>released"]
    B --> D["Kairos 7B<br/>released"]
    F["Jev and Laya<br/>System One decision approach"] --> E
    C --> E["Kairos v2<br/>in development"]
    D --> E
    B -.-> E

    classDef done fill:#0d3b66,stroke:#2e9ef7,color:#ffffff;
    classDef wip fill:#5a189a,stroke:#c77dff,color:#ffffff;
    classDef method fill:#1b4332,stroke:#52b788,color:#ffffff;
    class C,D done;
    class E wip;
    class B method;

Published on Hugging Face under the profile kirtzh.

• Papers in progress: CIDA, Calibrated methods, Multi-Agent systems
• Participation in conferences, with a growing record of talks and submissions

Every stage below is something I trained, broke and rebuilt, not just read about.

flowchart LR
    S1["Classical ML and NLP<br/>boosting, SVM, TF-IDF, LDA"] --> S2["Deep learning<br/>CNN, RNN, LSTM, Transformers"]
    S2 --> S3["Generative models<br/>VAE, GAN, diffusion"]
    S3 --> S4["Modern backbones<br/>U-Net, ModernNet, ResNet, EfficientNet"]
    S4 --> S5["LLM internals<br/>JEPA, sub-quadratic models,<br/>Jev and Laya"]
    S5 --> S6["Research<br/>CIDA, papers, conferences"]
    S6 --> S7["Open-source<br/>Kairos 1B, 7B, v2"]

    classDef a fill:#0d3b66,stroke:#2e9ef7,color:#ffffff;
    classDef b fill:#5a189a,stroke:#c77dff,color:#ffffff;
    class S1,S2,S3,S4 a;
    class S5,S6,S7 b;

I lead my university team in the ICPC. Competitive programming gave me a strong base in algorithms, data structures, graph theory, dynamic programming and complexity analysis, and it shapes how I write efficient training and inference code. Working in a team under time pressure also trained the habits I use in research: fast hypothesis testing, clean implementation and careful verification.