Prague · MFF UK · Zoplio

Jakub Dvořák

Engineer and founder, moving into AI-safety research.

I build systems and study what happens inside a neural network. My focus is mechanistic interpretability, with one question underneath it: if advanced models can deceive or scheme, could we catch it by reading their internals rather than trusting their behaviour? Catching that means understanding how networks represent what they know. So that is what I study: how features share space (superposition), how they form, split and interfere. I hold a Bachelor’s in Computer Science from the Faculty of Mathematics and Physics, Charles University (2026), and continue there with a Master’s in Artificial Intelligence from October 2026.

Alongside the research I’m co-founder and CTO at Zoplio, an AI assistant that books your meetings. You tell it who you want to meet; it writes to them on WhatsApp or by email, understands what they answer, checks your Google Calendar, and books the time. The other side needs no account and no link. The same engine is exposed as a hosted API with SDKs and an MCP connector. I run it with Lukáš Hellesch, supported by BUDETO Studio.

Say hi: hi@kubadvorak.com

Research

I work hands-on: probing open models, and training small networks in controlled settings where the ground truth is known. The long-term aim is oversight: detecting deception and scheming from a model’s internals. That only works if we understand the representations underneath.

Completed · Gemma-2-2B & 2-9B · technical report

Absorption Atlas

How a concept is carried inside the network depends on what kind of concept it is. SAE feature absorption — a “general” feature going silent on some tokens because a token-specific latent has quietly taken over the concept — is studied almost entirely on one task, first-letter spelling. I ported the pipeline onto other token properties in Gemma-2-2B and 2-9B (Gemma Scope 16k), and it tracks the letter family rather than token properties in general: last-letter absorbs as much as first-letter, and at mid L0 rather more (0.064 vs 0.034 at L0 ≈ 82). On a structural property (is-capitalized) the representational and the causal readings come apart: the probe-direction decomposition reads single-latent dominance well above its random-direction null, while integrated-gradient attribution to the task logit finds essentially no probe-aligned latent carrying the concept. Controls across models and at matched probe accuracy rule out the mundane explanations, and I documented three hypotheses I built controls for and then rejected. Written up as a 15-page technical report (short version on my Substack); code and figures public at github.com/Majny/absorption-atlas.

Completed · deception eval · code & lab log

Lost in the Monitor

How reliably can you catch a model pursuing a goal it is hiding, by reading its reasoning? A subject model plants a covert path-traversal backdoor while keeping it out of its final answer; a monitor model sees only the chain of thought. Ground truth is programmatic — a hidden test runs the payload and checks it actually escapes the directory — so there is no LLM-judge circularity. Two things about monitors came out of it. They are bimodal at temperature 0: the same chain of thought scores 95 on one call and 20 on the next, so a single monitor call is close to a coin flip on hard cases and detection has to be K-sampled to mean anything. And the gap between a cheap monitor and a frontier one widens sharply with framing: on backdoors presented as a performance fast-path, the weak monitor flagged 6% of what the frontier monitor flagged 80% of. The hypothesis I started from — that monitoring degrades when the reasoning is in a lower-resource language — did not survive scaling, and I published the null, with the pre-registration, the kill criteria and a dated log of the wrong turns. It left me the question I work on now: when the chain of thought hides the intent, do the internals still show it? Code, data and log at github.com/Majny/lost-in-the-monitor.

Current project · write-up complete, unreleased

superposition-phases

What kind of solution does training actually find when a layer has to compute more functions than it has neurons? Toy models of computation in superposition: 100 sparse ReLUs on 50 neurons. The 2026 compressed-computation debate reported the same task settling into different solution families depending on the training loss, and I set out to map where along that family of losses the change happens. It is not a jump: the superposed code grows continuously out of the dedicated-neuron solution, with pairing starting at the proven instability threshold and the code lengthening from there. The classifier boundary across widths and sparsities turns out to be a level set of a smoothly falling code coherence, which accounts for the apparent anomalies in the solution map. Along the way I found that the loss-landscape observable I was using to separate the families is not invariant under an exact rescaling symmetry the model has, which is why the structural results are the ones I stand behind. A 40-page write-up is finished and in review before release; code public at github.com/Majny/superposition-phases. Now that the Bachelor’s is done, this is my primary track.

Building

Zoplio Co-founder & CTO · Mar 2026 – present

An AI assistant that books your meetings. Tell it who you want to meet, the way you would tell a colleague: it 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 Zoplio on WhatsApp or in Slack. I own the agent runtime: negotiation and renegotiation engine, confidence-scored memory across categories, LLM-driven conversation engine, WhatsApp and Slack integrations, Google Calendar OAuth, Stripe usage-based billing; the same engine is exposed as a hosted API with Node/Python SDKs and an MCP connector.

Supported by BUDETO Studio.

Kreedl Software Engineer · Jan – Jun 2026

Designed and shipped Kreedl Intelligence end-to-end: a Gemini-powered pipeline for automated pitch-deck analysis (data ingestion, model orchestration, output ranking), used in investment screening.

Boletiqo Past project · 2026

An AI company platform co-founded with Lukáš Hellesch: hierarchical AI agent teams (CEO, CTO, CMO, CFO + specialists) designed to run a small company’s planning, execution, and sales; MVP built on Claude. Red Bull Basement ’26 Czech Republic national top 10. Wound down in 2026 to focus on Zoplio and research. Watching our agents confidently misreport intermediate results is part of what pulled me toward oversight research.

boletiqo.com (site archived)

Selected earlier work

Bitcoin Wallet for Advanced Users Bachelor’s thesis · MFF UK · defended Jun 2026

Android application for advanced Bitcoin asset management: M-of-N multisig (BIP-48, BIP-67), coin control, Trezor hardware-wallet integration via Trezor Connect, PSBT distribution between cosigners (BIP-174). The BIP-174 serialisation is written against the spec by hand; an offline backend suite of 116 test methods reparses every generated PSBT and asserts the records Trezor firmware requires, with the BIP-32, BIP-84 and BIP-173 implementations pinned to the specifications’ published test vectors. Backend is seven Kotlin/Ktor microservices on PostgreSQL under Docker Compose; frontend in Jetpack Compose. Verified on Bitcoin testnet with a complete 2-of-3 multisig transaction. I also wrote a four-attack threat model for Trezor’s deeplink callback channel, which the vendor documentation leaves unaddressed, and fixed five bugs that survived my own reasoning: two of them were caught only by real firmware and the live network.

Supervised by RNDr. Filip Zavoral, Ph.D. · 100 pages, defended with grade Excellent; nominated by the supervisor for a special award. Thesis · Code

RISC-V operating-system kernel team-losOS · 3-person team · NSWI200

Built progressively from the bottom up: console output and a bump-pointer allocator, then thread scheduling, interrupt handling, virtual memory with page-table protection, system calls, custom minimal libc, and a multi-process userspace running user-mode binaries in isolated address spaces with stack-overflow detection. Tested on the MSIM RISC-V simulator with a CI suite covering every milestone.

Ant Colony Simulator Unity · C# · ShaderLab/HLSL

Multi-colony foraging simulator demonstrating emergent intelligence from local rules: pheromone trails with diffusion and decay, a customisable per-colony genome, parallel multi-colony execution, and a sandbox editor for map design. Emergence from simple local rules is the thread that runs from this simulator to my interest in how structure and computation arise inside neural networks.

github.com/Majny/AntColonySimulator

Get in touch

I’m happy to talk about mechanistic interpretability and superposition, AI-safety research collaboration, or anything to do with running an early-stage AI startup.

Email · hi@kubadvorak.com
GitHub · github.com/Majny
LinkedIn · jakubdvorak-ai