AntJanusAntonin Januska

About

Hi, I'm Antonin.

Staff Engineer building AI-backed systems, engineering standards, and the workflows that let teams ship serious software.

Portrait of Antonin against greenery

Summary

The short version

Staff engineer focused on the infrastructure production LLM systems run on: evaluation harnesses, multi-provider dispatch, rate-limit handling, and guardrails on model output. Twenty years of building software, most of it lately in domains where a confidently wrong answer is worse than no answer.

LinkedInGitHub
20+Years building
2012Blogging since
600+GitHub stars

Selected work

Principles

Five proverbs I actually work by

  1. A deployed MVP is worth two prototyped.
  2. Dance like nobody is watching, code like everybody is.
  3. Today's fashion is tomorrow's legacy.
  4. Debugging becomes easier if you first admit that you are the problem.
  5. The best code is no code at all.

See the proverbs project →

Antonin hiking in Colorado

Beyond code

The rest of the operating system

Photography
I shoot animals, landscapes, people, and street scenes. The camera is how I stay observant.
Fiction
I write AI-themed fiction and short stories when technical writing is not enough.
Podcasts & Games
I have opinions about Final Fantasy, narrative design, and the overlap between games and technology.

Capabilities

What I work with now

The stack changes. The throughline doesn't. I'm usually working where backend systems, product delivery, and AI workflows collide.

AI & LLM Systems

Measuring model output, routing it across providers, and trusting none of it by default.

LLM Evaluation & Judge GradingMulti-Provider LLM DispatchToken Budgeting & Rate-Limit EngineeringStructured Output & GuardrailsModel Context Protocol (MCP)Agent Tooling & Orchestration

Languages

Python on the backend, TypeScript for everything else, SQL wherever the data lives.

PythonTypeScriptSQL

Platforms & Tools

Async services on Kubernetes, databases picked per job, front ends in React or Svelte.

FastAPIRedisPostgresSQLiteKubernetes on AzureReact & Next.jsSvelteKit

Leadership

Setting technical direction and standards, then reviewing what ships against them.

Technical StrategyEngineering StandardsCross-Team DeliveryCode ReviewMentoring

Contact

Let's work together

I help teams implement AI-enabled product work, sharpen engineering practices, and build systems that survive real usage.

Get in touch →

Ways we can work

  1. AI Workflow Audit. Typical scope: 1-2 weeks.
  2. Implementation Sprint. Typical scope: 2-6 weeks.
  3. Leadership Advisory. Typical scope: monthly advisory.

Your reading desk

Reading settings

No single setting suits every reader. The defaults come from reading research; change anything and the whole site follows. Your choices stay in this browser.

I build with AI and write about what I learn. Twenty years of software, written down so the next person has an easier time.

Text size

The British Dyslexia Association recommends 16 to 19px body text. The default here is 20px.

Line spacing

1.5 is the style-guide floor. Extra space between lines keeps your eye on the right line.

Letter spacing

Wider letter spacing reduces crowding. In one study, children with dyslexia read about 20% faster with it and made half as many errors (Zorzi et al., 2012).

Line length

Characters per line in the main text column.

Typeface

Plain, upright sans serifs did best in an eye-tracking study of 48 readers with dyslexia (Rello & Baeza-Yates, 2013).

Page tint

Pure white can glare. Pick the tint that is easiest on your eyes. Tints apply to the light theme.

Theme
Reading ruler

A tinted band that follows your pointer, like a card held under the line you're reading.