I am a senior C++ and Python engineer working where GPU performance, quantum software, and reliable developer tooling meet. My background spans Linux and embedded systems, production APIs and cloud platforms, CUDA experiments, and current open-source work across Qiskit, Qibolab, and NVIDIA CUDA-Q.
I like engineering that can be inspected and reproduced: explicit assumptions, measured results, focused tests, documented trade-offs, and commands that work for the next developer.
Much of my recent work uses coding agents for implementation and review, with every change personally validated; the merged contributions below provide the audit trail.
A reproducible systems proof of concept that runs the text portion of a packed 95.2 GB NVFP4 checkpoint on a 32 GB RTX 5090 by keeping most weights in pinned host memory and staging them through reusable native C++ NIXL/UCX requests.
- A byte-verified 504 MB request reached 28.407 GB/s through NIXL/UCX,
matching the 28.210 GB/s direct
cudaMemcpyAsynccontrol on the reference host. - A native double-buffered pipeline overlapped UCX transfers with CUDA work and measured 1.276x over the sequential estimate.
- The functional 128B run generated deterministic output at 0.3765 token/s mean end to end; the repository reports this as proof of operation, not a production serving result.
- Published artifacts record exact revisions, commands, host/build provenance, full-buffer verification, raw samples, and UCX lane selection.
An auditable, offline-first proof of concept for comparing quantum workloads across transparent synthetic device profiles. It combines deterministic Qiskit transpilation, typed agent orchestration with LangGraph and Pydantic AI, local evidence retrieval over SQLite FTS5, guarded noise simulation, and reproducible reports.
- Topology-aware placement produced 19.7% fewer two-qubit gates for the square-grid profile than the heavy-hex-style profile on a 13-case golden set.
- A documented native CUDA 13 /
sm_120Qiskit Aer build produced 20/20 CPU/GPU fidelity agreements within 2.22e-16. - The repository includes its Qiskit Aer source patch, build provenance, committed evaluation artifacts, CI, and a five-minute demonstration path.
The device profiles and costs are deliberately synthetic. The project separates measured compiler/simulator behavior from illustrative planning assumptions instead of presenting proxies as hardware results.
The five merged contributions below landed in June 2026; two began during unitaryHACK 2026.
| Project | Contribution and evidence | Status |
|---|---|---|
| Qibolab | Added GPU execution through JAX, CuPy, QuTiP, and Dynamiqs, with API controls, performance and accuracy benchmarks, tests, documentation, and reproducible dependencies. | Merged |
| Qiskit C++ | Added parameterized-circuit support to QASM3 export, including production code, tests, and release notes. | Merged |
| NVIDIA CUDA-Q | Extended mx and my measurement operations over qubit ranges with C++ API and integration-test changes. |
In review |
| Dynamiqs | Added a higher-order automatic-differentiation mode for Hessian computation. | Merged |
| Amazon Braket SDK | Added interactive Plotly circuit visualization to the Python SDK. | Merged |
| rustworkx | Updated the CI Clippy and rustfmt toolchain and carried the change through review. | Merged |
- Distributed CUDA-Q experiment notes — an illustrated field note on a successful 32-qubit statevector run across two Vast.ai nodes and four GPUs. It records the topology and CUDA-aware Open MPI networking lessons rather than presenting a reusable codebase.
- Local AI service — FastAPI and React with an OpenAI-compatible Ollama backend, retrieval, validation, observability, tests, CI, containers, and TLS ingress.
- Czech GPT-2 — an early publicly released Czech GPT-2 model, fine-tuned on Czech Wikipedia on a single GTX 1080 Ti; model notes and provenance.
- Qiskit Advocate, 2025
- Qiskit Global Summer School mentor, 2026
- Czech native; fluent English used daily in international engineering teams
I am based in Czechia and work comfortably across time zones.