Nebulae
A Belgian software company focused on building intelligent document parsing tools and solutions for human-machine collaboration in industrial fields such as building automation.
Timeline (last 90 days)
Customers of the Belgian software company Nebulae initially used ChatGPT directly for engineering drawing IO list extraction, with an accuracy of about 50%.
The human-machine collaborative document parsing tool built by Nebulae raised engineering drawing IO list extraction accuracy to above 95%, and it is already used daily in real projects.
Nebulae's technical lead said that after using the tool, the time to process one drawing dropped from two hours to 'checked it before the coffee got cold.'
Nebulae uses a method of inputting two views per page of drawing: the original text layer guarantees label accuracy, and the rendered image preserves the subordinate relationships between symbols and surrounding elements.
Nebulae built its own preprocessing pipeline to extract the precise position of every element on the page, letting the model read the drawing in left-to-right, top-to-bottom reading order rather than in the random order stored internally in the PDF.
Nebulae requires the model to directly return a fixed-structure, ordered IO list, so what the engineer gets is a checkable finished framework.
Nebulae extracts few-shot examples from the customer's past projects such as heat pump installations, server room ventilation, and HVAC systems, letting the model learn that team's naming rules and device pairing methods.
Nebulae's tool provides a review interface for each extraction result: the original drawing is shown on the left, the AI's structured reading result on the right, and engineers can check and correct it within seconds.
Nebulae chose Google Gemini, on the grounds that it performs strongest on true multimodal reasoning, can read the rendered image and the underlying text simultaneously in a single pass, and is cheap enough.
Nebulae emphasizes that the model choice was a snapshot at the time rather than a recommendation, and models iterate very quickly.
1 more timeline events
All Facts (11)
Customers of the Belgian software company Nebulae initially used ChatGPT directly for engineering drawing IO list extraction, with an accuracy of about 50%.
50%UnverifiedThe human-machine collaborative document parsing tool built by Nebulae raised engineering drawing IO list extraction accuracy to above 95%, and it is already used daily in real projects.
50%UnverifiedNebulae's technical lead said that after using the tool, the time to process one drawing dropped from two hours to 'checked it before the coffee got cold.'
50%UnverifiedNebulae uses a method of inputting two views per page of drawing: the original text layer guarantees label accuracy, and the rendered image preserves the subordinate relationships between symbols and surrounding elements.
50%UnverifiedNebulae built its own preprocessing pipeline to extract the precise position of every element on the page, letting the model read the drawing in left-to-right, top-to-bottom reading order rather than in the random order stored internally in the PDF.
50%UnverifiedNebulae requires the model to directly return a fixed-structure, ordered IO list, so what the engineer gets is a checkable finished framework.
50%UnverifiedNebulae extracts few-shot examples from the customer's past projects such as heat pump installations, server room ventilation, and HVAC systems, letting the model learn that team's naming rules and device pairing methods.
50%UnverifiedNebulae's tool provides a review interface for each extraction result: the original drawing is shown on the left, the AI's structured reading result on the right, and engineers can check and correct it within seconds.
50%UnverifiedNebulae chose Google Gemini, on the grounds that it performs strongest on true multimodal reasoning, can read the rendered image and the underlying text simultaneously in a single pass, and is cheap enough.
50%UnverifiedNebulae emphasizes that the model choice was a snapshot at the time rather than a recommendation, and models iterate very quickly.
50%UnverifiedNebulae explicitly cautions: these accuracy numbers come from one customer's drawings, and must be measured yourself first on different drawings; any percentage should be verified on your own documents before being trusted.
50%