Prague · MFF UK · Zoplio
Engineer and founder, moving into AI-safety research.
I build systems, and I 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’m finishing my Bachelor’s at the Faculty of Mathematics and Physics, Charles University (thesis defended in June, state exam this September), and start a Master’s in Artificial Intelligence there this autumn.
On the side, I’m co-founder and CTO at Zoplio — an AI scheduling primitive — with Lukáš Hellesch. We are backed by BUDETO Studio.
Say hi — hi@kubadvorak.com
I want to understand how neural networks represent what they know — how features share space (superposition), how they form, split and interfere. 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 — which only works if we understand the representations underneath.
A self-directed study of superposition in the wild that became an
audit of a field-standard interpretability metric.
SAE feature absorption — where a “general”
feature silently fails on some tokens while a token-specific latent
quietly carries the concept — is measured in the literature on
essentially one task: first-letter spelling. I ran the two standard
absorption metrics on a structurally different property
(is-capitalized) across Gemma-2-2B and 2-9B and found
they disagree: the current SAEBench-standard projection
metric reports modest single-latent absorption where the older
causal metric reports a clean zero — and cross-model and
matched-accuracy controls rule out the boring explanations. The
honest headline is a metric-validity caveat: a representational
absorption signal with no single-latent causal correlate. I also
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.
The AI scheduling primitive: a hosted API and SDKs (Node, Python, Anthropic MCP) any application embeds to delegate meeting negotiation. I own the agent runtime — agent-to-agent negotiation protocol, confidence-scored memory across categories, LLM-driven conversation engine, WhatsApp Cloud integration, Google Calendar OAuth, Stripe usage-based billing.
Backed by BUDETO Studio (€50k pre-seed).
An AI company platform co-founded with Lukáš Hellesch — autonomous AI agent teams (CEO, CTO, CMO, CFO + specialists) that plan, ship, sell, and close. 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.
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). Backend implemented as Kotlin/Ktor microservices on PostgreSQL with Docker Compose; frontend in Jetpack Compose. Verified on Bitcoin testnet with a complete 2-of-3 multisig transaction.
Supervised by RNDr. Filip Zavoral, Ph.D.
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 simulator with a CI suite covering every milestone.
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.
I’m happy to talk about mechanistic interpretability and the reliability of its tools, 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