AI Drawing

SPB Drawing Intelligence: from client drawings to an ERP-ready quote, with the estimator always in control

The problem. At SPB Metal, a sheet-metal and tube fabricator, every quote starts the same way. An estimator opens a pile of client files: PDFs, STEP models, DWG and DXF drawings, ZIP archives, often in Spanish or English. She reads every title block, parts list and dimension by hand. Then she works out the material, thickness, finish, blank size and cut lengths, and the operations each part needs, finds the matching codes in GPAC (our ERP), and types it all in. One project can mean dozens of parts and well over a hundred files. It is slow and repetitive, and a single misread thickness or wrong code becomes a wrong price.

What we built. A local web tool that reads the drawings for the estimator and prepares the quote in three steps: upload, review and save.

• Reads any format. It takes PDF, STEP, DWG, DXF and ZIP, pairs the 2D and 3D files of the same part, and shows a preview of each, including generated views of the STEP models.
• Extracts the technical data. From the title blocks, parts lists and geometry it pulls the part reference and name, material, thickness and finish (in Portuguese, Spanish or English). It also gets the flat-blank size, either measured or computed from the bends, and the tube profiles with their cut lengths.
• Matches real ERP codes, never invented ones. Materials are mapped to existing GPAC article codes using SPB's own rules, for example zinc-plated sheet uses the pickled base article plus an external zinc-plating line. When nothing fits, the line is created as a new article with a GPAC-style description. When several codes fit, the estimator chooses.
• Proposes the operations. It suggests laser sheet or tube cutting, bending, welding, threading, deburring and finishing, plus the fixed design, cleaning/packing and final-inspection steps. External services such as painting and plating are listed without times.
• Learns from the estimator. Each confirmed operation time is stored with the part's features. Once enough history exists, new suggestions are adjusted from the most similar past parts, and the tool shows which parts it compared.
• Exports for the ERP. It produces one Excel file per article, ready for GPAC import, and archives copies of the drawings in a folder per client. It then clears its workspace so each new job starts clean.

Human in the loop by design. Nothing goes to the ERP unverified. Every suggested time must be confirmed by hand. Every computed or estimated value is flagged "to confirm", and missing data is listed. Prices stay in GPAC. The AI does the reading and the routine work, and the estimator keeps the judgment.

Built on real work. This came from a real need on our shop floor and was trained and tested on real client batches. In the latest round it processed 122 client files into 45 articles, every one with its dimensions filled in. Each correction the estimator made became a general rule for all future drawings, and an automated test suite of more than 370 tests keeps those rules from breaking.

Impact. It cuts hours of manual reading and typing per quote to minutes of reviewing, with fewer transcription errors and faster answers to clients. Over time, the history of confirmed times turns individual know-how into a company asset. The approach works for any fabricator that quotes from technical drawings into an ERP.