Projects, Research, and Contributions.

A selection of my projects and research in Software Engineering, Machine Learning, and Data Systems, as well as hackathon and social impact work.

  • Apple: Post-Training Task Specialists with SFT & On-Policy RL

    Building the post-training pipeline for small task specialists: soft-verified-generation SFT per SERA, then on-policy RL with SDPO (self-distillation policy optimization).

    The task is automated issue resolution, so I also designed and shipped the production system those models run in (internal issues to PR-ready code diffs, end-to-end) plus a ModernBERT triage classifier fine-tuned on ~10K historical tickets (85% top-3 accuracy for component routing). The pipeline fine-tuning pipeline (soft-verified-generation SFT per SERA; on-policy RL with SDPO, self-distillation policy optimization) to replace frontier-model calls with a sub-dollar-per-fix specialist, targeting ~5x faster time-to-fix.

    Industry
    Post-Training
    Reinforcement Learning
    SFT
    SDPO
    LLMs

    Apple

  • Rare Candy Lab 🍬: Pokémon TCG AI Battle Agent (Kaggle Silver Medal)

    A from-scratch RL stack for "The Pokémon Company - PTCG AI Battle Challenge Simulation", one of the largest Kaggle competitions ever (6,892 teams).

    A 40M-parameter policy/value transformer trained by behaviour cloning on search-teacher data, then masked PPO with a KL anchor against a prioritized self-play league, then one LoRA specialist per deck co-trained on a frozen base and fused back in. Decks themselves evolved with MAP-Elites under a win-rate fitness with a worst-case floor.

    At match time: determinized tree search (PUCT) over a belief model of the opponent's deck, on 2 CPUs with no GPU and no network. Every change gated by a tiered eval battery with real power analysis. Result: silver medal, top 3% of the leaderboard.

    ML Project
    Reinforcement Learning
    PPO
    Self-Play
    Tree Search
    Kaggle

    Kaggle leaderboard: search "Rare Candy Lab 🍬"

  • Enhancing Intraoperative Registration with Neural Radiance Fields: An Exploration of Loss Functions Effects

    My Bachelor's thesis at Harvard Medical School: intraoperative brain registration using Neural Radiance Fields as differentiable, implicit representations of brain surface geometry, enabling direct optimization of camera positions via backpropagation.

    Includes a model-agnostic implementation of neural registration in the nerfstudio framework and a comprehensive analysis of loss functions (L1, L2, SSIM, NCC, MI) on registration accuracy, convergence, and stability; insights directly relevant to image-guided neurosurgery.

    Research
    Neural Radiance Fields
    Neural Registration
    Neurosurgery

    Link to the paper

  • "API for the Generated Reality": 4D Gaussian Splatting-based generation out of text and images

    Built during the "Project Lovable" Hackathon, this system transforms user text or images into dynamic 4D scenes (3D + time). We fine-tuned the Wan2.2 Video model to generate temporally stable, environment-representative videos, and then trained a 4D Gaussian model to reconstruct and render the evolving scene.

    The result is an interactive spatial experience where users can explore, walk through, and even edit their generated environments in real time. Essentially creating a personal holodeck powered by generative AI.

    References:
    * https://github.com/NVlabs/LongSplat
    * https://huggingface.co/Wan-AI/Wan2.2-Animate-14B

    ML Project
    Generative AI
    4D Gaussians
    Video Synthesis
    Wan2.2 Video
    Hackathon

    Link to the post

  • checkBay - Fact Checking NLP System

    With this application our team won a second place at the "TUM.ai AI4Good Hackathon". Our application CheckBay is designed to combat misinformation across all major social networks and beyond by cross-checking all publicly available sources and providing users with relevant ones.

    ML Project
    NLP
    Vector Search
    Hackathon

    Link to the post

  • AI4Good Workshops - Organizer & Lecturer

    AI4Good Workshop by TUM.ai is a two-fold program in a cooperation with the biggest university in Ghana — KNUST. Over the course of two weeks, participants were introduced to the basic principles of artificial intelligence, various application domains and how AI can be used to solve challenges in Ghana.
    For the second week, teams were formed to address a specific real-world problem. At the end of the workshop, the participating teams presented their results and achievements to a jury.

    AI4Good
    AI Education
    Lecturer
    Volunteering

    YouTube Playlist

  • TidyAI - Automating Data Cleaning

    As part of the automated agents hackathon, I designed, developed, and deployed a web application.

    The app focuses on automatically cleaning and exploring the data, as well as designing new features for seamless Machine Learning Development.

    You can test the application here: app.tidyai.tech

    Web Dev & LLM Agents
    LangChain
    GCP
    UX / UI
    Design
    Next.js
    Tailwind CSS
    chatGPT API

    Link to full project description (github, demo, presentation)

  • Portfolio Website

    This website! I built it to showcase my work, to learn Next.js + tailwind, and improve my ux/ui skills.

    Web Dev
    UX / UI
    Next.js
    Tailwind CSS

    podolskyi.io

  • be_certAIn - Career Recommendation Application

    The project involved developing an app that collects user information such as skills, hobbies, and free text input. The data is processed using (at that time) OpenAI's GPT-3 model (now, switched to GTP-3.5), which generates personalized career suggestions based on the input. The app provides explanations of suggested careers. The project aimed to offer a quick and personalized career test with a user-friendly interface, while ensuring data privacy and including a wide range of career options.

    Personal Project
    GPT
    Hackathon

    Link to the pitch deck

  • The Humanitarian Aid Hotline for Ukraine

    "The Humanitarian Aid Hotline" is a humanitarian project developed by a team of dedicated data scientists based in Munich, amidst the backdrop of the Ukrainian crisis. Functioning akin to an emergency hotline, this innovative platform seamlessly connects displaced individuals with suitable relief organizations, optimizing speed and efficiency. Leveraging technology and multilingual volunteers, refugee inquiries are promptly directed to appropriate aid networks. The ultimate vision is to open source the system's code, empowering the creation of similar call centers globally, thus redefining emergency response during crises.

    Software Project
    Data Engineering
    Backend Development
    Volunteering

    Link to the certificate and description

  • Practicum Report Paper: Evaluating Test Data Quality by Measuring Latent Space Coverage, Density, and Sparsity with Deep Neural Networks

    In this practicum we covered a novel method to assess the quality of test data within neural networks by examining Latent Space Coverage (LSC) and the density/sparsity of the samples generated in this space. Utilizing the Neural Coverage (NLC) fuzzing criterion in an existing fuzzing framework, we applied it to a self-trained LeNet-5 model on the widely-recognized German Traffic Sign Benchmark (GTSRB) dataset. Our findings illuminate the challenges and considerations in high-dimensional spaces when measuring sparsity and density of the latent space, as well as potential solutions for addressing computational resource issues in dealing with complex models.

    Research
    DNN Testing
    Fuzzing
    Latent Space Coverage
    Sparsity & Density
    GTSRB

    Link to the paper

  • Seminar Paper: Protecting the Grid from Wildfire through Data Analytics

    This seminar paper dives deep into the escalating threat of wildfires to power grid infrastructure and local communities, analyzing natural and man-made causes. The paper utilizes data-driven approaches to propose a comprehensive framework for wildfire prevention, early detection, and disaster relief strategies. We spotlight wildfire prediction techniques using physical simulations and real-world data, as well as the use of fault localization as an early warning instrument. The paper explores the role of sensors and data engineering in these processes, focusing on their use in transmission systems. Furthermore, we discuss proactive grid control as a crucial strategy in wildfire scenarios, all while offering an innovative tri-layer solution for serving shut-off areas with alternative power sources. A strong emphasis is placed on using remote sensing technologies like LiDAR for threat identification and establishing response strategies.

    Research
    Data Engineering
    Data Analytics
    Predictive Modeling

    Link to the paper