Mike Czerwinski

Twenty years building software. Now making AI systems answer for themselves.

Eleven years writing code, nine running delivery. Today I build and audit AI agent stacks: memory, decisions, locks, and the verification layer between a model and its consequences. The method is the one I apply to my own work: every claim carries its receipt.

Talk to me → LinkedIn DM

Who I am

The path here.

Eleven years hands-on development across e-commerce, media, and consumer web (PHP, JAVA, node.js, Spring, MySQL, AWS).

Nine years as Scrum Master and Agile lead across banks (BNP Paribas, Citi), automotive R&D (FEV Polska, delivering for Deutsche Post StreetScooter electric fleet and Hochtief Germany), travel-tech (Sabre, in a SAFe environment).

Now founder of three product lanes under one obsession: making opaque AI legible enough to ship without lying to the user. Content provenance (sentixx.net), quantitative trading research (ziom), custom-tailored personal AI (jessi).

Based in Poland. Working in English and Polish.

How I work

Operator diligence is the rare resource. Capability is cheap.

Five primitives I keep coming back to across every engagement. They are the parts most AI products skip until something breaks in production.

Auditable decisions with explicit lifecycle, not silent overwrites.
Defended locks on what must not move, enforced at admission, not at retrieval.
Source-attributed memory with per-atom provenance, not flat conversation history.
Write-time invariants that reject confident-but-unverified output before it propagates.
Refusal as first-class output. The model saying "I will not answer this" is a feature, not a failure mode.

I call this working frame agile4ai. More on it post by post on dev.to/jugeni.

Every number I publish carries the query it came from. Yours will too.

What I offer

One core engagement. Everything else follows from it.

Agent Stack Audit

Fixed scope, two weeks. I go into your agent system and examine memory architecture, decision lifecycle and coherence across sessions, lock semantics, source attribution, and the verification surface. On request: a multi-model red-team of the claims your product makes about itself.

You get a written report where every finding carries its evidence, and a debrief call. How it runs: one call to scope it, two weeks in the stack, report, then debrief. Engagements are invoiced B2B (EU VAT) via Sentixx.Net P.S.A. One engagement currently in motion.

Usually after an audit

Build agent infrastructure for regulated contexts. Typed memory with source attribution, decision lifecycle, write-time invariants, refusal as a first-class output.

Difficult second reader for AI product narratives. If the landing page is louder than the engineering, I notice. I tell you. Quietly.

Compliance audits, on request. EU AI Act Art. 50, DSA wave 2, GDPR DPIA, Directive 2024/1385, age-verification (UK OSA, EU AVMSD, DE JuSchG). The same map I run on my own company internally.

Why now

Capability is no longer the bottleneck.

AI agents are shipping into production faster than anyone can understand the systems they run. The marketing layer ships in week one. The receipt under it ships never.

This used to be a hobby concern. It is becoming the operating reality across regulated industries, finance, security-sensitive consumer software, and any product that delegates real decisions to a model.

The bottleneck is the layer between the model and the consequences: who verifies what, when, and on what evidence. That layer needs engineering, not opinion pieces.

Receipts

Public work, built systems, peer convergence.

Published — dev.to

Every Post I Publish Gets AI Review. A Hostile Agent Still Found the Holes in Twenty Minutes.
Red-teaming my own catalog: a review mandate is not an adversarial one, and I published the errata to prove it.
I Mined 2,505 Traders. The Only Edge Was What Not to Do.
Out-of-sample audit of 2,505 lead-trader signals on a major derivatives venue.
A published win rate is the actor auditing itself.
Funnel teardown of advertised vs. measured win rates across 9,312 Telegram signal messages.
They Said My Comment Was AI. So What?
Provenance as a verification problem, not a moral category.

Essays — Substack

Vibe Coding Has Two Axes. The Second One Doesn't Get Cheaper.
Cession and verification are independent axes. Better models raise cession for free; the check never gets cheaper.
Your Agent's Markdown File Is Becoming a Database With No Query Language.
Config layer versus memory layer: loading, precedence, and staleness once you pass the fourth markdown file.
Taste Is a Prediction That Survived Contact.
AI made fluency cheap. Taste is the predictor consequences trained, and it is what is left.

Built

sentixx.net
Polish simple joint-stock company. Content provenance and RegTech infrastructure (multi-tenant trust-and-safety engine).
jugeni
Agent infrastructure: typed memory, decision lifecycle, process threads, write-time invariants. Open source.

Peer convergence

kenielzep97: Your AI Obeys Rules That Expired. So Do You.
A practitioner running the staleness-and-authority frame against his own tooling, in public.
yongrean: Confidence is the one signal your model can't corroborate.
Took the published trading receipt as the canonical cross-domain example for his gate design.

Contact

Talk to me.

Email mike@mikeczerwinski.eu — engagement inquiries
Operational contact@sentixx.net — company matters
LinkedIn linkedin.com/in/mike-czerwinski-8784a752
Writing dev.to/jugeni
Essays jugeni.substack.com
Code github.com/jugeni

I read every message. I reply to the ones where the gap between what you have and what you need is engineering, not vibes.