Fifteen repositories, three games, four sites, three recordings and the music. Every description here was read from the repository itself rather than from its name — several of the earlier ones were wrong, and some of the repositories' own READMEs were wrong in the same way.
Published as evidence. The behavioural rules measured in the research did not start as research — these are two of the places they started.
34 packaged skills plus the framework they came from, in both .skill and plain markdown.
The rules measured in the research were not designed — they are survivors. This is where they came from, including the roughly 60% later pruned as waste. `custom-mode` enforced verification before a "done" claim a year before that became a measured rule.
github.com/misomalcek/claude-skills →A Tauri + Rust harness: a local/cloud model router, a vector layer, persistent context.
Where the same principles were first compiled rather than written — barycentrum.rs, ghost.rs, context.rs, snowglobe.rs. The direct ancestor of the current system.
github.com/misomalcek/atauri →Described from the README, the package description or the source — whichever was actually true.
A retrieval post-mortem: why a Cloudflare AI Search index returned nothing for a site whose pages were full of text, and what fixed it.
The same failure shape as three days spent blaming a ranking function — the answer was not in the index. Ships the worker, the server and the diagnosis.
open →mini-grokA native macOS chat window in Python against AppKit directly — no Electron, no web view, no framework. 352 lines.
open →auroraA multi-provider AI chat in one HTML file, 1,826 lines: streaming, live markdown, a collapsible reasoning panel, token stats, conversation memory, optional RAG grounding.
One file, several providers, no build step — an argument that a usable chat client does not need a framework.
open →gkggA Python wrapper around the Google Knowledge Graph Search API — 555 lines including its tests and a written build spec.
Entity grounding as one small tool that does one job and is tested.
open →genimA cyberpunk image generator for the hive — Express server against the Gemini image API, with the API notes it was built from.
open →hive-voice-echoA voice loop in Streamlit: record, transcribe with Google STT, answer, speak back with TTS.
The first end-to-end voice path here. It broke in the audio encoding and in language switching, not in the models — which is the part worth keeping.
open →gemma4-studioA minimal local studio for Gemma-4: one TypeScript server, one HTML page, no framework.
open →mini-harnessAgent definitions plus a written comparison of open-source multi-agent frameworks, done as research before committing to one.
The decision it produced was to build no framework at all — visible in the current system, which has none.
open →todoapiA REST API over a real SQLite database in 401 lines. Express, node:test, no ORM, 16 of 16 tests passing.
open →gsc-apiGoogle Search Console API scripts, rebuilt from a clean copy.
The original tracked a client's exported data — including in its history. Rebuilt rather than patched, because deleting a file does not remove it from git.
open →The values that used to sit on the homepage are not gone. They become a standalone portfolio project: a rules-axiom system with situational retrieval that a stranger can attach to their own AI. The repo is the next assignment, not this site.
Situational retrieval over behavioural rules — 1 in 5 by the rule's own wording, 10 of 10 by the situations that trigger it. Not a homepage values list. Source of the eight tested/untested items remains in this repo as data until that project ships.
Practical AI writing in Slovak — hands-on guides from work actually done, in a language where the alternative is translated marketing copy.
Source: gitlab.com/owlstar/aisvetvisit →The rebuild in progress: Astro, generated article imagery, new structure.
visit →An AI world-building experiment — generated worlds, rendered and navigable.
visit →A magazine-format publishing experiment.
visit →Most of my working life produced commercial projects under agreement — useful, measured, and not mine to publish. I never felt I had something a research community needed. That is different now: the AI work produces measurements, negative results included, and those belong in the open.
In every team I have argued for full sharing within approved eyes. I do not need to hoard what I know — I build on public knowledge myself, and anything properly researched and reviewed should end up back there.
Coursework, uploaded once around 2011 and never maintained since — and still accumulating readers, most of them from the United States, India, the United Kingdom, Malaysia and Germany. Verified live on 2026-08-20.
open profile →The Academia.edu profile is not here to suggest current academic activity, and it is no substitute for a Google Scholar record. It is here for two reasons: things I published once still resonate today, and I would very much like to publish again now that I have something worth publishing. It is also one of the sensible starting channels for that.
I financed my studies myself from nineteen, from the first day of my bachelor's, and finished with no student debt or loans. That sent me straight into business and practical application rather than research. The exception is the evidence: on an Erasmus stipend I did not have to work alongside studying, and that year I both enjoyed it and excelled at it — the same thing happening now with this essay and with AI research that is, objectively, still very limited.
My theoretical and academic complement to the explorer work — the AI systems knowledge corpus I built before we developed our own.
Thank you, Google, for supporting free public knowledge sharing — not without the mighty Google, ha? These are set to public. Opening one still requires a Google account.
The largest of them, and the one furthest from engineering. Dialogical principle in art and anthropology, quantum relationality, artificial life — read alongside agent orchestration and RAG papers in the same notebook.
Alterity became a value in the system's own axis: treat the other — human or model — as a being, never an object.
arXiv preprints on interactive computation next to epistemology-of-intelligence analysis and the internet read as a complex adaptive system.
Where "complex adaptive system" stopped being a phrase and started being the thing we were actually building.
Half external, half my own working documents: context engineering reports, programmatic tool calling, and the Content-DNA framework from the glossary build.
The SelfMoA note in here became a skill I actually use; Content-DNA is the framework that made an agency proposal obsolete.
Everything published on driving a real browser from inside an agent — testing, conversion of UI tests to API tests, the MCP skill implementations.
This is the instrument I verify with. Every measurement in this CV — the 17px floor, the mobile pass — came out of that toolchain.
Symbolic logic and truth trees, SHACL validation under ontologies, symbolic pattern planning, neuro-symbolic verification of LLM outputs.
The most recent, and the open direction: where a deterministic layer should sit under a probabilistic one.
Logic-of-Thought and LOGIC-LM++, evaluation metrics split into reference-based, reference-free and LLM-based, and Prolog revived as soft-coded logic.
Contains a neuro-symbolic blueprint for Gemma-4 with TypeScript as the rigid scaffold — which is close to what the local agent became.
Predicate logic translated back into English via dynamic semantics, self-evolving agents, semantic code retrieval — beside Logos, Gnosis and Nous as three capacities of the soul.
The pairing is deliberate: vector search is a theory of meaning whether or not you write one down.
Anthropic's own material on coordination patterns and on decoupling the brain from the hands, the agent harness as an infrastructure layer, and satellite anomaly detection as a working case.
Read before deciding NOT to build orchestration. Knowing the five patterns is what made removing ours a decision rather than a gap.
One dense source on orchestrator-worker models, routing, the Agent SDK and skill templates — with a Slovak audio overview generated from it.
The earliest of the set, and the plainest: this is where the agent work started as reading.
Systems running, rather than described.