About me - Krzysztof Sopyła, PhD

I am Krzysztof Sopyła — AI tech lead, agent systems builder, and concept-reasoning researcher.

Build agents by day, train models by night, publish the mess.

That is the working agreement I have with myself, and it is also the most honest description of what this blog is for.

During the day I lead AI teams at Pearson. I work closely with product owners, architect agent systems, bridge the gap between what stakeholders expect and what is technically feasible, and try to build a team culture where experimentation is safe and honest communication is the default. Over the years I have held roles from AI researcher to CTO in startups, to Chief AI Architect at my own software house (Ermlab). Each taught me something different: academia showed me rigorous thinking, startups taught me speed, and enterprise taught me that aligning people is harder than aligning models.

In the evenings I do my own ML research. I believe that leading AI teams without staying hands-on is a recipe for losing touch with reality. So I keep training models — and writing up the failures along with the wins.

What I’m Building #

Agent Patterns Lab — 9 design patterns for production multi-agent systems, with running code for each one. The progression goes from a single LangGraph pipeline to a cloud-deployed, authenticated, multi-team agent system using MCP and A2A. I built it because the gap between framework demos and enterprise-grade agent architecture is enormous, and almost nobody writes about the structural decisions that matter.

MrCogito — Concept Reasoning Model — open research on concept-bottleneck architectures: compressing sequences into a small bank of dense concept embeddings and reasoning over them. The vision is an efficient reasoning model at ~21M parameters — proving that architecture choices matter more than parameter count. Phase 1 ran into concept collapse and taught me a lot about what does not work. Phase 2 is now active and built around an agentic auto-research loop: the same agents I build by day plan and run the experiments I cannot run by hand at night.

What I Write About #

This blog covers three things:

1. Enterprise agent architecture — MCP, A2A, multi-agent patterns, and the gap between what agent frameworks promise and what production actually requires.
2. AI research in the open — honest documentation of MrCogito experiments, including the ones that failed and why.
3. Leading AI teams — stakeholder alignment, team culture, and the tension between innovation pressure and realistic delivery.

I don’t write hype. I don’t write marketing copy. I write about what I tried, what I built, what broke, and what I think it means.

Background #

My path through AI: PhD in computer science (ML focus, GPU-accelerated sparse matrix multiplication (KMLib library), SVM and online learning), then co-author of one of the first Polish BERT models, work on context-extended transformers for book understanding, recommendation engines with LSTMs, and a Polish writing assistant that started using Transformers back in 2018. I like revisiting old ideas with modern engineering — sometimes the future lies in better understanding the past.

I believe in open science, open source, and accessible education. I believe that even small, well-designed experiments can point you in the right direction. And I believe that sharing negative results is more valuable than pretending everything works on the first try.

You can find me on GitHub, LinkedIn, X, and Google Scholar. I used to write about AI in Polish on ksopyla.com.