KERVION

1. Architecture

I am developing a custom AI architecture that uses LLMs in a responsible, non-agentic way. This approach makes it possible to achieve near-zero operational errors by keeping language interpretation separate from action execution.

Sensorium transforms natural language commands into predictable actions executed by a deterministic system. The LLM serves solely as a perceptual layer: it extracts structured signals from language but does not make operational decisions.

Execution logic is separated into typed adapters, arbitration layers and deterministic executors. Language-derived signals are validated, normalized, and resolved before any action is performed. This reduces the impact of hallucinations and ambiguity, ensuring control, auditability, and reliability in business applications.

Multi-sensor LLM pipeline for controlled signal extraction● Typed adapters, arbitration layers, and deterministic execution
● Multilingual architecture with auditable control flow
● Rule-based operation resolution and state control
● Reliability-focused pipeline and end-to-end testing

Python, FastAPI, Vue 3, TypeScript, LLMs: OpenAI, Qwen

Possible applications: CRM, ERP, document management, customer support, internal tools, business process automation, transactional systems, an other business applications where control, auditability, and reliability are critical.

Menu is a demo of a transactional drive-thru application. It is an example of the Sensorium architecture, in which LLM/AI serves as the sensory layer.

● speech-to-text based on Google API
● LLM from OpenAI
● the demo also runs on small local LLM models

The purpose of the demo is to show that it is possible to build error-free, auditable services that use AI responsibly. The Menu architecture is original and is not a copy of any known design pattern.

2. Experimental AI Movies

AI fatigue, anti-hype, social backlash, collective layoffs, the paradox of automatic artText from: "The Scream Programmer"
Tools: ElevenLabs, Leonardo, Adobe, Vidu
LipSync: local free open source tools
Automation: n8n

3. LLM Training

I am currently creating a series of fine-tuned LLM models trained on publicly available texts by Polish poets and writers in the public domain, meaning authors who died before 1956, as well as on my own texts and styles.
Each author should have their own model. The base model is the Polish PLLuM. The base will be changed as soon as a more suitable model appears. The creators of PLLuM state: “We ensure that our model is safe, free from harmful and false content,” which makes it only moderately adequate for literature that reaches into shadow and darkness.

At the request of contemporary authors, I also create paid models based on their own writing.

Technical stack: PLLuM, LoRA, RunPod.

In connection with Olga Tokarczuk’s recent remark, “Darling, how beautifully we could develop this?” — we think similarly, except that in my case it more often turns into: “What the hell did you write here, you idiot?” Jokes aside: LLMs and neural networks are remarkable assistants in programming, art, literature, science, mathematics, and business. We have been given a powerful tool, and it will take years before we learn how to use it for the benefit of ourselves and others.

The question of authorship of works created with the use of AI remains open. Some guiding questions: who is the author of a novel written in a library? Which matters more: the author or the work? Is the purpose of art wonder, or ego? And what is the purpose of science? If you write beautifully, to whom do you owe the beauty of your style?

4. Attention Is All You Need

I am currently studying and translating into Polish the most famous paper of recent years: Attention Is All You Need.
(Niestety, nie mogę go w tej chwili tutaj opublikować ze względu na prawa autorskie — sprawa w toku.)

One of the co-authors is Polish: Łukasz Kaiser. I enjoy listening to interviews with him because he preserves a subtlety that is rare in the Google/OpenAI environment, both in style and in meaning. His approach is very close to mine. I will mention two key points from his interview with Maciej Kawecki:

LLMs do not think. They “think” in quotation marks.
LLMs are excellent assistants.

I follow the same path: assistant, not agent; simulation of thinking, not thinking; algorithm, not subject.

One of the ideas presented in the paper is multi-head attention. I see in it a kinship with the mechanism I use in Sensorium. I use linguistic multi-sensors that are triggered in parallel over the whole utterance. The goal is different, but the linguistic intuition is identical.

Tłumaczenia automatyczne

Na przykładzie tytułu chcę wskazać na ograniczenia automatycznych tłumaczeń i samych transformerów. Angielski tytuł Attention Is All You Need zostaje przerzucony do polskiego Uwaga to wszystko, czego potrzebujesz. Uważam, że polski tytuł powinien brzmieć: Wystarczą powiązania. Rdzeniem artykułu jest bowiem myśl, że kolejnym słowom wypowiedzi nie należy się przyglądać w kolejności ich występowania i na tej podstawie wektoryzować tokeny, ale znajdować ich powiązania w stosunku do wszystkich innych słów kontekstu. Termin „uwaga” brzmi po polsku niezrozumiale i koślawo. Polacy nie gęsi, powtórnie.

I drugi przykład Programista krzyku został automatycznie przerzucony do angielskiego Screaming Programmer. Poprawny tytuł angielski brzmi: The Scream Programmer. Transformer nie znajduje żadnego przykładu takiego wyrażenia, więc domyśla się błędnie, że programista krzyczał, a nie że programuje krzyk.

5. Automation

I build process automations, including AI-based processes, using n8n. The films above are one example. Another simple business use case is email triage using tools such as Airtable.

Programming automation tools from the big three — OpenAI, Anthropic, and Google — such as Codex, Claude, and Copilot, are still very immature as production tools. The situation may change tomorrow, but for now I do not use them on a daily basis. The number of errors made by agents and the high cost of tokens are not economically justified compared with cheaper, more specialized, and faster assistants.

6. Archives & Tools