~/romanzdk $ whoami

Roman Žydyk

Practical AI. Reliable Python. Data platforms.

I’m a technical lead with a data engineering background. I build data platforms in Python and work with AI agents: coding agents in daily team work, LLM checks with evals, and tools that keep people in control of what agents do.

SELECT * FROM ai_engineering

AI engineering

At work I help drive AI adoption across the company, together with the CTO: I look for new tools and approaches, try them on real work, and bring the ones that hold up to the team.

at work

Coding agents in a team

A workflow from request to merge request: a person approves the plan, tests run, the change is checked on the running service, and two agents review it. Built on shared skills and MCP integrations.

at work

LLM checks with evals

Smoke checks for internal APIs where a Pydantic AI judge reads live responses. An eval suite tests the judges on known-good and known-bad cases.

at work · my initiative

A company-wide agent kit

I designed and built a shared kit of agent skills and rules, an onboarding site generated from the same files, and a sandbox where an agent takes a task from chat. A working version, proposed for company-wide use.

at work

Testing an agent platform

I ran real, low-risk tickets through Paperclip, an agent orchestration platform, reported what broke, and contributed skills and fixes to the company setup.

side projects

ai-newsletter: a daily AI digest that checks its sources

2026 · open source · Python · Pydantic AI · Braintrust
problem
A plain summary of feeds repeats one event five times, sells old projects as new and drops the small print.
how
Plain code collects the sources. A model picks the events and checks each one against its source. It may read a few more pages when a key fact is missing.
measured
Evals on made-up cases compare it with a single-pass version: 0.92 against 0.70 on keeping the key condition. One run costs 2 to 3 cents.
links
Code · Notes from the evals
The model is used only where judgement is needed.
verifair – review the evidence, not the diffPython · FastAPI · htmx · Postgres

A self-hosted service where coding agents publish change reports with their evidence. People sign, comment and approve. Agents do not merge. Docs.

Personal assistant

in use

An assistant I use myself, with long-term memory, tools and recurring tasks: email drafts, source-linked digests, voice queries. Built from open-source parts such as Hermes Agent, Hindsight and Home Assistant.

How I approach it, with examples →

SELECT * FROM data_engineering

Data engineering

at work

Data platform

Design and operation of a Python data platform on Airflow, Kubernetes, Apache Iceberg and S3.

at work

Market-data APIs

Led delivery of equity and futures data APIs. One shared API and a Python client library replaced direct provider calls and repeated queries.

at work

Data quality and monitoring

Continuous API checks, one framework for historical data-quality checks, and alerts with less noise.

at work

Technical leadership

Technical direction, architecture decisions, mentoring and hiring, while staying hands-on in Python.

Experience includes Quantlane and PwC. MSc in Information Systems and Technologies, Prague University of Economics and Business, with a thesis on MLOps with Kubeflow. Full history on LinkedIn →

side projects

Live Czech election results

2021 · 2023 · 2025 · solo project
problem
I wanted clean, fast election-night results without ads and commentary.
how
A Python backend polls the official results feed, stores it in PostgreSQL and shows it in a Streamlit app. In 2025 I wrote the backend and let a coding agent generate the frontend.
links
Backend · Frontend · Story (Czech)
Architecture of the 2023 and 2025 versions.
S3 web browserPython · Flask · AWS

A simple web UI for browsing S3 storage and sharing files through time-limited links. Related: S3 Commander, a two-pane browser with server-side move and copy.

Older projects (2)

Winning team project at the Data Science & BI Academy (VŠE, 2019/2020). We merged 12 years of police accident data and compared classifiers. Interview (Czech).

13 public dashboards, from a decade of Forbes billionaires to table tennis in Czechia.

SELECT area, tools FROM skills

Skills

AI engineering
coding agentsPydantic AIMCPLLM evalsagent review workflows
Data engineering & platforms
SQLAirflowpandasPolarsArrow / ParquetApache IcebergPostgreSQLTimescaleDBRedisRabbitMQ
Python & software quality
FastAPIasynciotypingtestinglintingpackagingcode review
Cloud & operations
KubernetesDockerArgo CDGitLab CIAWSAzurePrometheusGrafanaSentryKibana
Technical leadership
technical directionarchitecture decisions1:1s & mentoringhiring & onboarding

SELECT title, format FROM teaching

Teaching & community

AI talks for the team

internal talks

Several talks for my team about using AI agents in everyday engineering work.

Rust for Pythonistasfree course · English

A free interactive course that teaches Python developers to read Rust. Every example is run in CI. Code.

Python: Code qualityonline course · Czech

Writing Python that is typed, tested, documented, secure and fast enough.

Database & SQLlive workshops · Czech

Relational databases and SQL, taught live for Turing Institut (formerly BeeIT).

Azure cloud basicsonline course · Czech

Cloud concepts and a practical tour of Azure services.

Tableau in practiceonline course · Czech

Using Tableau for data visualization in everyday work.

Česko.Digitalvolunteering

Data engineer and consultant on projects that helped people after the start of the war in Ukraine, such as Movapp and Stojíme za Ukrajinou.

SELECT email FROM contact

Contact

Interested in AI engineering, Python or data platforms? Get in touch.