Curriculum Vitae

Jakub Dvořák · Curriculum Vitae

Two printable variants below; pick by audience. For the long story behind the bullet points, see About.

Profile

Engineer and founder moving into mechanistic interpretability, aimed at detecting deception and scheming in AI systems. Bachelor’s in Computer Science from MFF UK (2026); starting a Master’s in Artificial Intelligence there in October 2026. Current focus: how neural networks represent what they know (superposition and feature geometry), studied by probing open models and training small networks where the ground truth is known, as the groundwork for oversight. Two completed research projects: Absorption Atlas, on how sparse-autoencoder features carry different kinds of token property, and Lost in the Monitor, on how reliably a hidden goal can be caught by reading a model’s chain of thought (both below). Co-founder and CTO at Zoplio (supported by BUDETO Studio).

Research

Completed 2026

Absorption Atlas: how SAE features carry different kinds of concept

First completed research project: a cross-property, cross-scale study of SAE feature absorption, a superposition phenomenon, written up as a 15-page technical report. Absorption is measured in the literature almost entirely on first-letter spelling; ported onto other token properties in Gemma-2-2B and 2-9B, it tracks the letter family rather than token properties in general. Last-letter absorbs as much as first-letter (0.064 vs 0.034 at L0 ≈ 82); on a structural property (is-capitalized) the representational and causal readings come apart, with the probe-direction decomposition well above its random-direction null where integrated-gradient attribution finds essentially nothing. Cross-model and matched-accuracy controls rule out task-ability and dead-layer artifacts. I also documented three hypotheses the controls forced me to reject. Cross-checked against published SAEBench results. Public repo: github.com/Majny/absorption-atlas; short write-up on Substack.

Completed 2026

Lost in the Monitor: a published null on cross-lingual CoT monitoring

A control-style deception eval with programmatic ground truth: a subject model plants a covert path-traversal backdoor while hiding it from its final answer; a monitor model sees only the chain of thought. The hypothesis (monitoring degrades in lower-resource reasoning languages) died at scale and I published the null, with the pre-registration, kill criteria and a dated lab log including the wrong turns. Two methodology results survived: monitors are bimodal at temperature 0, so single-call scores are unreliable and detection must be K-sampled; and monitor strength interacts sharply with subtlety (6% vs 80% detection on subtly-framed backdoors). Public repo: github.com/Majny/lost-in-the-monitor.

Current focus

superposition-phases: what training finds when a layer is overloaded

The live project: toy models of computation in superposition (100 sparse ReLUs on 50 neurons), mapping which solution family training finds across the family of training losses. The superposed code grows continuously out of the dedicated-neuron solution, with pairing starting at the proven instability threshold; the classifier boundary across widths and sparsities is a level set of a smoothly falling code coherence. A 40-page write-up is finished and in review before release; code public: github.com/Majny/superposition-phases. Laptop + GPU-cluster scale.

Education

Oct 2026 – 2028

Master’s in Artificial Intelligence

MFF UK, Charles University, Prague. Starting October 2026.

2023 – 2026

Bachelor of Science, Computer Science

MFF UK, 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, nominated by the supervisor for a special award; degree completed September 2026. Thesis (100 pp) · Code

2018 – 2023

Earlier

VŠCHT Prague (transferred, 2022–2023). Gymnázium Mladá Boleslav (2018–2022).

Experience

Mar 2026 – present

Co-founder & CTO · Zoplio

An AI assistant that books your meetings: you say who you want to meet, it writes to them on WhatsApp or by email, understands their replies, checks your Google Calendar, and books the time (renegotiating when someone’s availability changes); the other side needs no account and no link. I own the agent runtime: negotiation and renegotiation engine, confidence-scored memory across categories, LLM-driven conversation engine, integrations for WhatsApp, Slack, Google Calendar, and Stripe; the same engine is exposed as a hosted API with SDKs and an MCP connector. Supported by BUDETO Studio.

Jan 2026 – Jun 2026

Software Engineer · Kreedl

End-to-end design and implementation of Kreedl Intelligence, an AI-powered pipeline for automated pitch-deck analysis: data ingestion, model orchestration, output ranking.

Selected projects

2026

Boletiqo

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 to focus on Zoplio and research.

Bachelor’s thesis

Bitcoin Wallet for Advanced Users

Android application for advanced Bitcoin asset management: M-of-N multisig (BIP-48, BIP-67), coin control, Trezor integration via Trezor Connect, PSBT (BIP-174) hand-serialised against the spec. Backend: seven Kotlin/Ktor microservices on PostgreSQL with Docker Compose; frontend: Jetpack Compose. 116 offline backend test methods, with the BIP-32/84/173 implementations pinned to the specifications’ published test vectors. Includes a four-attack threat model for Trezor’s deeplink callback channel, which the vendor documentation leaves unaddressed. Verified on Bitcoin testnet with 2-of-3 multisig. Thesis · Code

team-losOS

RISC-V operating-system kernel

Built bottom-up: console output, bump-pointer allocator, thread scheduling, interrupt handling, virtual memory with page-table protection, system calls, custom minimal libc, multi-process userspace with stack-overflow detection. Tested on the MSIM RISC-V simulator with a per-milestone CI suite.

Unity · C#

Ant Colony Simulator

Multi-colony foraging simulator with pheromone trails, per-colony genome system enabling distinct “strains”, parallel execution, and a sandbox map editor.

Highlights

Red Bull Basement ’26 · Czech Republic National Top 10
SCIO Mathematics · 98th percentile
Podnikni to! · CTU Entrepreneurship Program

Skills

Languages: Python, TypeScript, Kotlin, C++ (modern), C, C#, Java, Bash.

AI / LLM engineering (shipped): LLM APIs in production (Claude, Gemini, GPT), Anthropic MCP, multi-agent orchestration, RAG, structured output.

Research toolkit (applied in shipped projects): PyTorch forward/backward hooks, TransformerLens, SAE tooling (SAELens · Gemma Scope), toy-model training & probing, SAE analysis on Gemma-2-2B/9B, local inference of open-weight models.

Frameworks: Node.js, Next.js, React, FastAPI, Ktor, Jetpack Compose, SFML.

Infrastructure: Docker, PostgreSQL, MongoDB, Redis, Stripe, Google Cloud, AWS, GitLab CI/CD.

Systems: Linux (Arch), RISC-V kernel development, Bitcoin protocol (BIP-48 / 67 / 174 PSBT, multisig), Trezor Connect.