Sensorium is a custom, non-agentic AI architecture for transforming natural language commands into controlled, predictable system actions.

The core principle is simple: language interpretation and action execution are kept separate. Large language models are used only as a perceptual layer, extracting structured signals from user input. They do not make operational decisions and do not execute actions directly.

The execution layer is deterministic. Signals extracted from language are validated, normalized, arbitrated, and resolved through typed adapters, rule-based operation mapping, controlled state flow, and predefined executors.

Invalid or ambiguous requests are blocked. In uncertain situations, the system prefers inaction over unsafe action.

This architecture reduces the operational impact of hallucinations, ambiguous phrasing, and incorrect model interpretations. It is designed for systems where reliability, auditability, and control matter more than autonomous behavior.

Demo: Sensorium

The current demo presents a deliberately simple set of visual operations: showing, hiding, selecting, highlighting, arranging, aligning, and resetting objects based on natural language commands.

The purpose is not visual complexity, but proof of controlled execution: the system performs only allowed operations, under explicit rules, with a traceable state flow.

Sensorium is applicable to business systems where natural language interfaces must remain safe and predictable: CRM, ERP, document management, customer support, internal tools, process automation, transactional systems, and other environments where language can help users operate software without giving the model authority over the system.

Technology: Python, FastAPI, Vue 3, TypeScript, OpenAI API, Qwen, ChatGPT, Ollama

Menu is a demo drive-thru ordering application and a practical business use case for the Sensorium architecture. It shows how a language model can serve as a sensory layer for ordering, while the actual transaction remains controlled, validated, and auditable.

The system accepts orders by voice or typed text and currently works in three languages: English, Spanish, and Polish. Voice input is supported through speech-to-text, with a daily demo limit of 30 voice commands.

Demo: Menu

The purpose of the demo is to show that AI can be used responsibly in a transactional environment. The model helps interpret natural language, but order validation, ambiguity handling, negation control, and action safeguards remain part of the deterministic application logic.

This makes the system suitable for scenarios where mistakes are not acceptable: food ordering, customer service flows, internal tools, transactional interfaces, and other business processes where natural language must be translated into precise, verifiable operations.

Technology: Google Speech-to-Text API, OpenAI, Qwen, Python, React, FastAPI, TypeScript

Sensorium is not an ordinary use of LLMs. They can be used in many different ways, but today they are most often expected to “think”: to draw conclusions, interpret, decide, and provide ready-made answers. Here, the purpose is different. LLMs do not serve as a mind, but as sensors: instruments that register subtle patterns of meaning, tension, and similarity.

This is not a popular approach in the contemporary use of language models, but its advantage is fundamental: it makes it possible to reduce almost to zero the errors that arise when LLMs are entrusted with tasks for which they are not the best tool.

A broader philosophical description of this method has been published elsewhere. It has not yet been translated into English.

The approach described above is inspired by a philosophical tradition that includes, among others:

Ludwig Wittgenstein — Philosophical Investigations
Kurt Gödel — On Formally Undecidable Propositions of Principia Mathematica
Alfred Tarski — The Concept of Truth in Formalized Languages
Łukasz Kaiser — Attention Is All You Need (co-author)
Gary Marcus — Taming Silicon Valley: How We Can Ensure That AI Works for Us
Karen Hao — Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI
Alan Turing — the halting problem

In various ways, this approach is also related to the perspectives of critics who run YouTube channels:

Julian Whatley
Errol Brandt / Kiraa
Mo Bitar
El / House of El
Yann LeCun
Parth Patel / ParthKnowsAI

We do not include here a long list of well-known names — our intellectual adversaries — whose views on LLMs and the human being depart significantly from the sober approach represented by the people listed above.