About
A Czech computer scientist who ships production systems and is turning that discipline toward AI safety.
I’m a Czech computer scientist working at the intersection of applied AI and AI safety: shipping production systems as a founder while moving into mechanistic interpretability: the mathematics of what actually happens inside a neural network. This page is the long version; the printable one is the CV, and the Now page tracks what I’m focused on today.
Jakub Dvořák is a Czech software engineer based in Prague, moving into AI-safety research. He is co-founder and CTO of Zoplio (supported by BUDETO Studio), an AI assistant that books meetings: you tell it who to meet on WhatsApp or in Slack, and it writes to them on WhatsApp or by email, checks your Google Calendar, and books the time. He holds a Bachelor’s in Computer Science from the Faculty of Mathematics and Physics, Charles University (MFF UK, 2026), where he begins a Master’s in Artificial Intelligence in October 2026. His research focus is mechanistic interpretability (superposition and feature geometry) as groundwork for detecting deception and scheming in AI systems by reading their internals rather than trusting their behaviour. His first technical report, Absorption Atlas, measures how sparse-autoencoder features carry different kinds of token property in Gemma-2-2B and 2-9B, and finds that absorption tracks a family of properties rather than being a universal failure mode. A second, Lost in the Monitor, asks how reliably a model pursuing a hidden goal can be caught by reading its chain of thought. His current project, superposition-phases, maps which solution families training finds in toy models of computation in superposition. He writes at kubadvorak.substack.com and publishes code at github.com/Majny.
I started programming as a teenager and went through the eight-year selective gymnasium track at Gymnázium Mladá Boleslav. I briefly studied pharmaceutical synthesis at VŠCHT before transferring to computer science at the Faculty of Mathematics and Physics, Charles University (MFF UK), where I completed my Bachelor’s degree in 2026.
My early projects were systems-flavoured: a RISC-V operating system kernel with a 3-person team, a custom Bitcoin transaction backtester in C++, an Ant Colony Simulator in Unity exploring emergent intelligence from local rules. The thread connecting them is curiosity about what makes complex behaviour appear from simple components.
That same thread pulled me toward AI safety, and specifically toward the question of whether we could catch a model that is deceiving or scheming by reading its internals rather than trusting its words. That is a bet on mechanistic interpretability, and it only pays off if we actually understand how networks represent what they know: how features share space (superposition), how they form, split and interfere. That is where I work: probing open models, and training small networks in controlled settings where the ground truth is known. My first completed research project, Absorption Atlas, measures how differently a network carries different kinds of concept; the second, Lost in the Monitor, asks how much of a hidden goal is legible in a model’s reasoning at all. The full story of both is on the home page. Understand the representations first; then use them to look for deception.
My focus is mechanistic interpretability, aimed at a concrete safety target: detecting deception and scheming by inspecting a model’s internals. The near-term work is upstream of that: understanding how networks represent what they know, and when the standard readings of those representations mislead. It plays to my comparative advantage (the linear-algebra and statistics of representations) and it runs on a laptop plus a GPU cluster.
Deliberate convergence, not a new direction every month: production AI → oversight → the representations oversight depends on → superposition and how to read it. Lost in the Monitor asked how much of a hidden goal is legible in a model’s words, and pointed at the internals; Absorption Atlas measured how the internals carry a concept in the first place; superposition-phases asks what structure training puts there when the layer is overloaded.
Research is now my primary focus; I remain co-founder and CTO at Zoplio, handing day-to-day work to the team. We’re building an AI assistant that books your meetings. You tell it who you want to meet, the way you would tell a colleague, and it does the rest: writes to them on WhatsApp or by email, understands what they answer, checks your Google Calendar, and books the time, renegotiating with everyone affected when someone’s availability changes. The other side needs no account, no app, and no link; you talk to it on WhatsApp or in Slack, and it works in English, Czech, Slovak, or German. The technical core is an agent runtime: a negotiation and renegotiation engine, confidence-scored memory of how you actually work, an LLM-driven conversation engine, and integrations for WhatsApp, Slack, Google Calendar, and Stripe; the same engine is exposed as a hosted API with SDKs and an MCP connector. We’re supported by BUDETO Studio.
I also co-founded Boletiqo with Lukáš Hellesch: an AI company platform of hierarchical agent teams designed to run a small company’s planning, execution, and sales. We reached Red Bull Basement ’26 Czech Republic national top 10, and I wound it down in 2026 to focus. Watching our own agents misrepresent intermediate results and produce convincing post-hoc justifications at the edge of their capability was a formative push toward oversight research.
I view the founder side as a useful counterweight to research: it forces me to take AI deployment realities seriously, and it puts me in the room when models actually break.
And one free-time side project: I run GazeUp: daily AI news in Czech (news, research, startups), with short-form video on Instagram and TikTok, everything with cited sources.
Starting October 2026 at the Faculty of Mathematics and Physics, Charles University. Research theme: mechanistic interpretability: superposition and how networks represent concepts, aimed at detecting deception and scheming in AI systems.
Specialization: Systems Programming. Bachelor’s thesis: Bitcoin Wallet for Advanced Users, supervised by RNDr. Filip Zavoral, Ph.D., defended 18 Jun 2026 with grade Excellent and nominated by the supervisor for a special award; degree completed September 2026. Thesis (100 pp) · Code
Brief detour at the University of Chemistry and Technology Prague (VŠCHT, 2022–2023, transferred). Selective gymnasium at Gymnázium Mladá Boleslav (maturita 2022).
I keep these lists separate on purpose. The first is what I’ve been paid to ship; the second and third are shipped but unpaid: the thesis wallet, a RISC-V kernel, a Bitcoin backtester. The last is the research toolkit I’ve applied in Absorption Atlas and Lost in the Monitor, and am deepening with every project.