Federico Cerbelli
Automated Information Extraction from Technical Drawings Using Vision-Language Models: A Human-in-the-Loop Approach.
Rel. Paolo Garza. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
Abstract
Before a fastener can be ordered, engineered or approved, its characteristics must be read from the technical drawing supplied by its manufacturer and typed by hand into a central database. Repeated across thousands of components and hundreds of suppliers, each with its own drafting conventions, notations and table structures, this manual transcription is slow, error-prone and difficult to scale. This internship, carried out at FORVIA within the Group Data Office & Digital Transformation, addressed that problem through the design, implementation and evaluation of an end-to-end drawing-extraction system built entirely on the Palantir Foundry platform. The system reads drawings with a vision-language model that receives, alongside the target document, an annotated reference drawing acting as a one-shot visual example and a high- resolution crop of the title block.
Extracted values never enter trusted data directly: they are surfaced in an application in which engineers review, correct and approve them
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
