AI and backend engineer in Prague, trained in systems programming. I build backends and LLM systems, and I'm available now for a full-time role or a contract. BSc Computer Science, Systems Programming, MFF UK (2026) · Master's in AI there since October 2026.
kubadvorak.com · CV · hi@kubadvorak.com
-
Zoplio developer API (source private, surface public): API docs ·
@zoplio/sdk-js·zoplioon PyPI ·@zoplio/claude-tool. Zoplio is an AI assistant that books meetings over WhatsApp, Slack and email. As co-founder and CTO (Mar to Oct 2026) I wrote all of its backend, including:- the public API: keys, rate limits, usage quotas and signed webhooks;
- both SDKs (Node/TypeScript and Python);
- the hosted MCP server, so AI agents can book meetings.
In October 2026 we put Zoplio on hold: on its own, it was a feature rather than a standalone product.
-
bitcoin-multisig-wallet: my Bachelor's thesis (100 pages on DSpace, write-up), graded Excellent. An Android M-of-N multisig wallet in which every signature is made on a Trezor via Trezor Connect; the backend coordinates the cosigners but never holds key material and never signs.
- Hand-written BIP-174 (PSBT) encoding and parsing, the multisig P2WSH script and the bech32 encoder.
- 7 Kotlin/Ktor services on PostgreSQL under Docker Compose.
- 123 backend test methods in GitHub Actions.
- Runs any policy from 2-of-3 up to 15-of-15 on testnet (the 15-of-15 on the Trezor emulator); also used on mainnet with real funds.
-
Kreedl Intelligence (private; AI Engineer at Kreedl, full-time, Jan to Jun 2026): I designed and shipped an internal Gemini pipeline for pitch-deck analysis in investment screening. Python, FastAPI, MongoDB, Docker.
-
sectionsight: an agent skill for Claude Code, Codex and Gemini CLI that sets up cookieless website analytics, measuring which sections people actually read, and checks the result in a real browser with Playwright.
-
Systems programming at MFF UK (my Bachelor's specialization; course code, not public):
- a RISC-V (Sv32) teaching kernel in C, built in a team of three: heap, scheduling, synchronization, virtual memory, user-mode processes;
- a C-subset compiler to LLVM IR (flex, bison, IRBuilder);
- parallel programming with TBB, OpenMP, SYCL (CPU target), Spark and C# async, with jobs on the faculty's SLURM cluster;
- Linux Administration: QEMU/KVM VMs as systemd services, DNS with DNSSEC, backups verified by a restore.
- PostHog/posthog-js #5325 (merged): session replay kept recording network requests after the client had turned that off; fixed, with unit and Playwright tests.
- Documentation fixes merged in trezor/trezor-suite and sebbo2002/ical-generator.
With Lukáš Hellesch: I write the code, he does the product side.
- AirGap Mandates, FinTechers Hackathon (Oct 2026): a bank's AI agent proposes payments but can't pay; money moves only within a mandate signed on a Trezor, and every step goes into a log anyone can verify offline.
- Heirloom, Cursor Hackathon Prague (Sep 2026): an AI grandson that phones Grandpa and turns his stories into a family archive.
- Grant Ops, Prague Builder Day, Duvo × Anthropic (Jun 2026): an agent that checks a research grant's expenses against the funder's rulebook.
- Boletiqo, Red Bull Basement '26: Czech national top 10.
What ties it together is measurement: checking whether the numbers we use to read a model actually hold up. I'm heading toward the overlap of high-performance deep learning and AI safety.
- absorption-atlas: how a concept is carried inside Gemma-2-2B/9B depends on what kind of concept it is. Built with TransformerLens, SAELens and Gemma Scope, in GPU jobs on MetaCentrum; 16-page technical report.
- superposition-phases: toy models of computation in superposition. The public code is the preliminary first pass; the 256-job CPU sweep and the 40-page write-up are not public yet.
- Memory in long-context LLMs: does a model know which part of its KV cache it will need before it needs it? Running; first measurements through November.
- Earlier: lost-in-the-monitor, a published null result on chain-of-thought monitoring, with an eval harness around gpt-oss-20b under vLLM.
Day to day I build with Claude Code: I write a short spec first, keep each change small and read every diff before it goes in. The thesis wallet and my kernel code I wrote almost entirely by hand.