Marcin J. Ołdak

— 01 · About

I ship working things.

Not slides, not past tense — live demos you can try on this page.

I've been exploring AI since 2022 and turned it into my profession in 2026 with MJ.OLDAK — my AI consulting practice. Out of dozens of experimental agents, I brought 4 products to production-ready for small businesses: invoice OCR, an email assistant, a tax-deadline agent and a report generator.

My daily toolkit is LLMs in practice — Claude (API + Claude Code), Gemini — and n8n workflow orchestration, including a production pipeline exposed as an MCP tool for AI agents.

Lately I've gone deeper on the data side — SQL, reproducible analytics pipelines and source-cited reporting on public datasets (see the analysis below).

My background is unusual on purpose: business psychology sharpens discovery and designing around the user's mental model; law makes GDPR, human-in-the-loop and ethical AI second nature — designed in from day one, not bolted on.

Since 2022 I've also worked inside a large international corporation — a 100% English-speaking, heavily regulated environment where I automate repetitive work bottom-up.

production-ready AI products

2022

working with AI since

3.7y

English-only corporate env

— 02 · Work

Don't read it — try it.

Four AI products I built for small businesses, as interactive demos.

Everything below runs in your browser on sample data — no sign-ups, no tracking. The production versions run on FastAPI + Claude API and are deployment-ready for Polish accounting offices.

Live · production

Sales Agent B2B — live demo

A real, deployed app: form → GenAI lead classification (HOT/WARM/COLD with reasoning) → database with row-level security. GDPR Art. 22 compliant.

Open demo.mjoldak.pl ↗

Agentic infrastructure

n8n pipeline as an MCP tool

Production n8n workflow (Webhook → Gemini → Supabase EU Frankfurt), green end-to-end and exposed via the Model Context Protocol — so AI agents can call it as a tool.

Webhook → LLM → Parse → DB (EU) · instance-level MCP

Problem

Single-task automations break the moment reality gets messy — and nobody sees why.

Solution

An autonomous agent that plans, calls tools, recovers from failures — and shows its reasoning at every step.

Effect

End-to-end process handled. Supervised, not hands-off: a human makes the final call.

Agent Mission Control

Watch an AI agent work

Pipeline

○

Webhook intake

n8n · webhook

○

OCR extraction

Azure DI + Claude

○

Validate tax ID

checksum + registry

○

Post to ledger

Supabase EU

○

Draft confirmation

Claude

○

Human approval

you · the gate that ships

Agent log — live

$ Pick a scenario and hit „Run agent" — you'll see the agent plan, call tools, hit a real-world failure and recover from it. Right here in your browser.

Scripted simulation mirroring my production architecture (n8n Webhook → LLM extraction → validation → Supabase EU → human approval) — including how agents should fail and recover. No external calls; runs entirely in your browser.

Problem

Accounting teams retype invoice data by hand — minutes per document, errors included.

Solution

PDF or photo in → structured data out (Azure Document Intelligence + Claude API).

Effect

An invoice processed in seconds, ready for the ledger.

OCR Invoice Reader

Interactive demo · sample data

Document

FAKTURA VAT

BUDMAT Hurtownia Materiałów Budowlanych Sp. z o.o.

ul. Składowa 18, 53-127 Wrocław

NIP: 899-284-71-63

Data wystawienia: 04.06.2026

Pozycje: cement portlandzki, wełna mineralna…

Extracted fields

Pick an invoice and hit „Process invoice" — extraction runs right here in your browser.

Production version: FastAPI + Azure Document Intelligence + Claude API — PDF/photo in, structured data out. The scanned document is intentionally a Polish invoice — that's what the product reads. Demo runs entirely in your browser.

Problem

Client emails bury accountants — urgent tax-office letters mixed with spam.

Solution

Classification into 5 categories + a drafted reply for each; a human approves before anything is sent.

Effect

Inbox triage on autopilot, replies in minutes.

Email Assistant

Interactive demo · sample data

Inbox — click an email

Classification & reply draft

Open any email — the assistant classifies it and drafts a reply, right in your browser.

Production version: IMAP/SMTP + Claude — classification into 5 categories + reply drafts written in Polish for the client, human-in-the-loop by design. This demo runs in your browser on sample emails (in English for readability).

Problem

Monthly client reports mean hours of repetitive copy-paste work.

Solution

Ledger exports in → branded PDF report out, with KPIs and a written summary.

Effect

Minutes per client instead of hours.

Reporting Agent

Interactive demo · sample data

Client

Month

Pick a client and month, then generate — the monthly report builds right here in your browser.

Production version: FastAPI + openpyxl + fpdf2 — branded PDF reports per client (in Polish), generated from the office's ledger exports. This demo runs entirely in your browser on sample data.

— 03 · Skills

Stack & methods.

Tools I work with every day — not a logo wall.

Agentic AI & LLMs

▸ Claude (API + Claude Code) — daily
▸ Gemini
▸ autonomous agents
▸ prompt engineering
▸ multi-LLM routing, provider abstraction

Workflow orchestration

▸ n8n (workflow + MCP tool)
▸ Make.com
▸ Google Apps Script
▸ REST APIs & webhooks
▸ modern alternative to traditional RPA

RAG & knowledge bases

▸ RAG patterns in products
▸ knowledge-base design
▸ personal Zettelkasten — 270+ concepts
▸ AI-ecosystem monitoring (26 sources)

Discovery & business analysis

▸ Mom Test discovery
▸ BPMN 2.0 (AS-IS → TO-BE)
▸ user stories, Given/When/Then
▸ PRD, ROI modeling
▸ compose-vs-build framework

Ethical AI & compliance

▸ GDPR Art. 22 / 28
▸ human-in-the-loop by design
▸ EU data residency
▸ data privacy & security by default

Programming (AI-assisted)

▸ Python, FastAPI
▸ SQL, Supabase/PostgreSQL (RLS)
▸ Azure Document Intelligence
▸ Astro, Cloudflare Pages

Data analysis & pipelines

▸ SQL — JOIN, window functions, reconciliation
▸ DuckDB, pandas
▸ pandera data contracts
▸ DVC — reproducible pipelines
▸ data provenance (SHA-256)
▸ public data: GUS · Eurostat · NBP · ECB

— 04 · Open Source

Code in the open.

Don't take my word for it — read the source.

A few public repositories — agents, reusable tooling and reproducible analytics. Each card links straight to GitHub.

Python

agent-flow

Multi-step ReAct document agent with a human-approval gate and cited reports.

View on GitHub ↗

Claude Code

claude-skills

Reusable Claude Code skills toolkit: RAG-with-provenance, repo publish audit, case-study & workshop builders.

View on GitHub ↗

Python

analytics-stack-pl

Reproducible, zero-cost analytics pipeline (DVC · pandera · DuckDB · Quarto) on AI adoption in Polish vs EU enterprises.

View on GitHub ↗

Frameworks

ai-ba-decision-frameworks

Decision frameworks for the AI Business Analyst role (CC BY 4.0).

View on GitHub ↗

See all on GitHub →

— 05 · Data & Analytics

From data to evidence.

Public data → an audited pipeline → a report that survives scrutiny.

Beyond shipping agents, I do the analyst's core job: find the right public data, reconcile the same metric across independent official sources, benchmark and trend it — and make every figure traceable and reproducible. Here is one finished piece, end to end.

SQL · DuckDB · DVC · pandera Reproducible

AI adoption in Polish enterprises

A GUS × Eurostat reconciliation · 2021–2025

The same indicator — enterprises using any AI technology — reported by two statistical offices, reconciled to within 0.3 points; Poland is climbing fast off a low base, yet the gap to the EU average is widening. Economy-wide data, with the financial-sector caveat stated upfront.

0.34pp

two official sources reconciled (GUS vs Eurostat)

26→31

Poland's rank among EU countries, 2021→2025

−11.6pp

and the gap to the EU keeps widening

Poland vs EU27 AI adoption trajectory, 2021–2025 — Fig. 1 · Poland vs EU27 trajectory

AI adoption by firm size, 2025 — Fig. 2 · Adoption by firm size (2025)

What it demonstrates

▸SQL — JOIN · window functions · reconciliation
▸Reproducible pipeline (DVC) — one command rebuilds it
▸Data contracts (pandera) — fails loud on drift
▸Provenance — SHA-256 + dataset vintage on every source
▸Honest scoping — limits stated, not hidden

Read the report (PDF) ↗ View the pipeline ↗

AI & Automation Consultant

I ship working things.

Don't read it — try it.

Sales Agent B2B — live demo

n8n pipeline as an MCP tool

Stack & methods.

Agentic AI & LLMs

Workflow orchestration

RAG & knowledge bases

Discovery & business analysis

Ethical AI & compliance

Programming (AI-assisted)

Data analysis & pipelines

Code in the open.

agent-flow

claude-skills

analytics-stack-pl

ai-ba-decision-frameworks

From data to evidence.

AI adoption in Polish enterprises

I write about AI
for business.

Let's talk.

Marcin J. Ołdak

AI & Automation Consultant

I ship working things.

Don't read it — try it.

Sales Agent B2B — live demo

n8n pipeline as an MCP tool

Stack & methods.

Agentic AI & LLMs

Workflow orchestration

RAG & knowledge bases

Discovery & business analysis

Ethical AI & compliance

Programming (AI-assisted)

Data analysis & pipelines

Code in the open.

agent-flow

claude-skills

analytics-stack-pl

ai-ba-decision-frameworks

From data to evidence.

AI adoption in Polish enterprises

I write about AIfor business.

Let's talk.

I write about AI
for business.