About
Hey there! I am a Research Engineer at JetBrains, working on post-training models for agentic applications in software engineering. I studied computer science at the Technical University of Munich, where I completed my master’s degree with distinction in November 2025.
My work is broadly centered on software engineering agents (for code and beyond) that robustly solve tasks over long horizons, and adapt continually from experience. I am particularly interested in how we can make such agents efficient and effective—in terms of context management, memory, and compute—so that they are capable of robustly operating in complex real-world computer systems and remain accessible to a wide range of users and organizations.
My path into research was shaped by years of work as a full-stack engineer in long-term internships and data science roles. This foundation gives me both the technical depth to build robust research infrastructure and an intuitive understanding of the developer workflows and pain points that drive my research questions. It also taught me to communicate complex technical ideas across disciplines—bridging the gap between rigorous ML research and practical engineering challenges.
If my research interests or projects resonate with you, feel free to reach out or explore my projects and writing below.
- View selected work on the projects page.
- Check out my invited talks.
- Read my notes and write-ups on the blog.
Latest updates
Gave a guest lecture for TU Munich’s Advanced NLP Masters’ course on proactive context management for long-horizon code agents: moving from harness-level heuristic summarization to learned meta-actions, where the model itself recognizes stale tool output or a dead subtask branch and folds it away on its own initiative. Closed with a hands-on demo of using subagents to scale qualitative trajectory analysis for debugging long-horizon agent runs. See the talks page for more.
Invited talk at Berlin DataTalksClub: Context Engineering for Agentic Hybrid Applications with Ivan Potapov. We discuss the cost and reliability implications of context growth in long-horizon production agents, and argue for efficiency-focused, proactive context management. Building on findings from The Complexity Trap that simple observation masking matches LLM summarization at a fraction of the cost. Watch on YouTube.
I published a reproduction and improvement of the RouteLLM BERT router, identifying and fixing critical methodological flaws in the original implementation.
Completed my master’s degree in computer science at the Technical University of Munich with distinction.
The Complexity Trap was accepted to the Fourth Workshop on Deep Learning for Code (DL4C) at NeurIPS 2025. We show that simple observation masking is as efficient as LLM summarization for agent context management and introduce a novel context management strategy that reduces costs by up to 11% compared to these baselines.