R&D

Local AI as an engineer's assistant.

I develop AI tools that run on a local computer, in two directions: machinery compliance documentation and work on the shop floor. AI proposes; the engineer checks and approves.

Research problem

How can locally run AI models support the engineer in technical documentation and in the analysis of production processes, so that every AI proposal can be checked against the source and approved by a human?

Principles of the method

A language model works in a narrow, defined role. Requirements and calculations that follow from regulations and standards determine the rules written in code.
A control gate holds the document when critical data is missing.
Research direction: detecting where critical values disagree with the approved source, not only where data is missing.
I record quality thresholds before measuring.
An engineer approves the document.

Status scale

A status applies to a tool. A project does not inherit the status of its most mature tool.

Concept: A description of the solution; no working code, or a sketch.
Prototype: Working code, checked on synthetic data or on own examples.
Laboratory testing: Measurement on a frozen dataset with a threshold recorded before the measurement; the type of data is stated with the result.
Validation in real-world conditions: A pilot with a user, on real documents, with agreed indicators.

Machinery compliance documentation

CENTAUR

Prototype

A system in which the engineer prepares the entire compliance documentation of a machine in one place, from classification to declaration.

Works today

classification of the machine and its conformity assessment route
selection of standards
risk assessment with a hazard register
essential health and safety requirements of Annex III, with a verdict for each
client files (photos, PDFs, spreadsheets) attached as evidence, with detection of changes to the file
gates before the declaration (requirements, hazards, life phases, links between hazards and requirements, manufacturer data)
change log
declaration of conformity and documentation package
work on a local computer, without the cloud

In preparation

an AI assistant that proposes which requirements apply to the machine and points out inconsistencies (the engineer always makes the decision)
analysis of test results and recordings
references to regulations and standards at every step
data for the nameplate and a QR code for the declaration
several people working on one project, with a change history and one approving person

Limitations: The gates check completeness, not the correctness of values. The system covers Regulation (EU) 2023/1230 only.

Shop floor

Workstation instruction generator

Prototype

A workstation instruction built from what the worker shows and says.

Works today

approval by an engineer

In preparation

a draft instruction from a voice recording and photos of the workstation
control of critical values (tolerances, tightening torques, limit values) against the source documentation
a video recording as a source
loading an existing instruction or product documentation
language versions for multilingual teams (e.g. Ukrainian, Spanish)
a QR code for the workstation and for each language version, available from a phone next to the paper version

Workstation assistant

Concept

The workstation as a whole: layout, access to tools and parts, ergonomics, safety.

The image is processed on site. Before a pilot: a data protection impact assessment and agreements with the employer.

Direction of development

analysis of the workstation and the course of work with a vision system, to help the worker, not to evaluate them
step-by-step instructions on a screen at the workstation, for new employees
a shadow board on the worktop with the right tools and parts highlighted by colour
an incorrect step signalled immediately
connection with the workstation instruction generator

Production flow analysis

Concept

Direction: local analysis of work in progress, machine states and material flow from production data, without analysing employee behaviour.

Technology collaboration

I am looking for manufacturing plants and machinery manufacturers to validate the work on real documents. I share detailed test results and the methodology with R&D partners.

Let's talk about R&D collaboration