Industrial AI automation

Project Overview

NT Group, an industrial manufacturer of steel components, faced inefficiencies in manually identifying spare parts, causing delays and increased workload. They required a modern digital solution to swiftly and accurately identify components, optimizing their support processes.

Hatimeria, our partner firm, analyzed possible solutions and proposed integrating advanced AI and interactive 3D visualization technology, including 3D CAD, PLM data, and LiDAR scanning, into a seamless digital workflow.

AI Engineering

Tanner Lab was tasked with solving the problem of identifying the correct specification documentation from an image of a spare part (e.g., a customer-taken photo). Leveraging recent advancements in multimodal Deep Learning models trained on both images and text, we developed a Retrieval-Augmented Generation (RAG) proof of concept.

Background: What is RAG?

Retrieval-Augmented Generation (RAG) is an AI method combining retrieval (searching relevant context) and generative AI models. RAG first retrieves relevant information, then generates accurate, context-aware responses using this retrieved context.

Source: Turtlecrown (CC BY-SA 4.0)

Proof of Concept

Our PoC included:

• Node.js command-line tool integrated with a FastAPI API.
• CLIP large (AI model) to embed images and textual specifications into high-dimensional vectors.
• LanceDB vector database for efficient storage and retrieval.
• Retrieval performed via cosine similarity and k-nearest neighbor (kNN) search.

Hatimeria’s Further Development:

Hatimeria expanded our PoC into a comprehensive solution featuring:

• An interactive 3D configurator using three.js.
• Advanced AI chatbot capabilities powered by the Mistral model.
• Integration with CAD models (SOLIDWORKS), PLM data (Roima/Aton), and LiDAR scans.
• Full e-commerce integration for streamlined ordering.

Outcome: The collaboration resulted in a powerful digital system, significantly enhancing spare part identification accuracy and efficiency, reducing downtime, manual efforts, and overall operational costs.