Selected work

    Projects

    Some applied ML, agentic systems, and practical AI tools I've been working on.

    Gemma 4 Software-Engineering Agent · Current Focus

    Agent harness for the Google × Kaggle Developer Agent Competition that turns a quantized Gemma 4 31B model into a software-engineering agent resolving real GitHub issues from private repositories. The constraints drive the design: a 32k context window, a fixed evaluation tool set, and about 60 tasks per 12-hour run that scores zero if it overruns. Most of the work is context management, time budgeting, and extending the fixed tools with reusable skills and agent-written code execution, iterated through a logged experiment loop.

    LLM AgentsGemma 4Agent HarnessContext EngineeringEvaluationKaggle
    Repo coming soon

    Local-First AI Agent · Ongoing

    A local-first personal research agent running on my personal MacBook M3, using local LLMs via Ollama, a custom MCP server exposing tools (PDF reader, file search, writers), and a controller orchestrating LLM–tool interactions. Includes a small document-store RAG and an interactive CLI interface.

    LLMsAgentsMCPRAGCLI ToolsLocal Inference
    Repo coming soon

    Kaggle Notebook Master · Ongoing

    Kaggle Notebook Master focused on model development, feature engineering, and clear, educational notebooks. My notebooks have collected 2,900+ upvotes and 2,400+ forks, and I peaked at #75 worldwide in the notebooks ranking.

    XGBoostLightGBMCatBoostEDAML Competitions
    View my Kaggle profile

    StudentsAI · Paused

    Open-source software development project: an AI study toolkit with a FastAPI backend and structured data model for notes, flashcards, SRS, and activity analytics. Includes an AI service for summaries, flashcard generation, keyword extraction, and note graphing, with OAuth login, email verification, and rate-limited OpenAI integration. Currently being re-scoped into a mobile-first flashcards app.

    FastAPIPostgreSQLAuth/OAuthLLM APISRS FlashcardsNote GraphBackend Architecture
    View on GitHubLive demo

    Mimic Recognition

    Latest contribution to my AI-Playground repo, an educational collection of small ML/AI projects with hands-on notebooks and explainers. This project is a real-time facial expression recognition system, trained on FER2013 and fine-tuned on a custom dataset.

    Computer VisionDeep LearningCNNsPyTorch
    View on GitHubLive demo

    Micrograd Pro

    An enhanced, educational take on Karpathy's micrograd: a scalar autograd engine built around a Value class, extended with modern activations (GELU, LeakyReLU), LayerNorm, Dropout, and sequential layers. Interactive notebooks visualize gradient flow and computation graphs, with mathematical explanations and paper references throughout.

    AutogradBackpropagationNeural NetworksNumPyEducational
    View on GitHub

    © 2026 Oleksii Z.