Ebara Corporation Integrates On-Premises AI Solutions for Confidential Technology Management

Introduction

Ebara Corporation, a leading global industrial machinery manufacturer, has implemented the on-premises AI solution "GBase On-Prem" from Sparticle Corporation, headquartered in Chuo, Tokyo. This strategic integration addresses the complex challenge of leveraging generative AI in manufacturing while safeguarding highly confidential technical information. Currently, Ebara is utilizing GBase On-Prem for validating and operating its internal knowledge to enhance operational efficiency.

Background: Barriers to Cloud AI Adoption in Manufacturing

Ebara Corporation possesses a vast repository of sensitive information, including design documents, technical manuals, and internal regulations, developed over many years. While these resources secure Ebara’s competitive edge, transferring them outside the company has been strictly controlled.

The rapid proliferation of cloud-based generative AI technologies, such as ChatGPT, has raised various concerns for Ebara:

- Risk of sending confidential data to external servers
- Concerns around data sovereignty and security policies
- Precision challenges that arise from the cloud models failing to meet the specialized accuracy demanded in manufacturing tasks

Due to these factors, full-scale implementation of AI solutions in business operations remained a daunting hurdle for Ebara. An official from the Cyber Physical Strategy (CPS) department expressed the sentiment, "Generic AI systems fall short of providing the expert-level responses required in manufacturing. Yet, we cannot expose crucial technical information to external environments."

Selection Criteria: The Optimal Solution

Given these challenges, Ebara opted for the on-premises generative AI solution, GBase On-Prem. The primary considerations leading to this choice included:

1. Complete Data Sovereignty: Operating entirely within the internal network allows utilization of confidential data like technical sketches and corporate regulations without risking external exposure.
2. High Precision in Japanese Language Processing: Testing demonstrated that Ebara’s 70B parameter model achieved results comparable to GPT-4o mini, offering exceptional accuracy suitable for on-premises operations at a manageable cost.
3. Flexible Model Configuration and Quantization Support: Ebara can choose from models like 70B or 80B according to specific needs. The solution utilizes 8-bit quantization technology alongside NVIDIA A100 GPUs to effectively optimize GPU resources.
4. Customized RAG Capabilities: The Retrieval-Augmented Generation (RAG) mechanism enables the training of an Ebara-specific AI using internal documents, making this a decisive factor for its deployment.

Benefits: Enhanced Efficiency and Knowledge Utilization

GBase On-Prem is currently in the Proof of Concept (POC) phase, which has been running for about two months. Feedback from the users indicates a satisfaction score of 9 out of 10. Notably, the benefits realized include:

1. Streamlined Technical Information Retrieval

Previously, the average time for searching technical information was 30 minutes. With GBase On-Prem, users can now fetch information using natural language queries in just 2 minutes. Examples of time reduction include:

- Search for past design documents: 30 minutes to 2 minutes (93% reduction)
- Locating relevant sections in internal regulations: 20 minutes to 1 minute (95% reduction)
- Cross-referencing technical documents: 45 minutes to 3 minutes (93% reduction)

2. Ease of RAG Development

Users without technical expertise can now build a knowledge base simply by uploading documents, promoting a user-led approach to utilization.

3. Secure Testing Environment

The elimination of external data leak risks allows for AI testing with previously untouchable confidential data, enhancing overall data security.

Future Prospects: Advancing AI Solutions for Manufacturing

Sparticle is committed to further enhancing AI utilization support tailored to the manufacturing sector.

- Development Support for Industry-Specific Models: Increased focus on building fine-tuning models utilizing high-performance computing resources like H200 GPUs. This includes understanding industry-specific terms, design philosophies, and quality standards.
- Expansion of Visual Language Model (VLM) Support: Upcoming features will leverage visual data such as drawings and inspection photographs for comprehensive analytical functions. Combining text and image inputs promises a more intuitive information retrieval process.
- Promotion of Business Automation through Agent Functions: Planned feature expansions for GBase On-Prem include:

- Unified search spanning multiple systems
- Automated summarization and comparative analysis of multiple documents
- Automatic generation of standardized reports
- Detection of issues and improvement suggestions based on user logs

The Sparticle development team emphasizes their goal: "We aim to provide an environment where our manufacturing clients can maximize their substantial technical assets using AI. Transforming AI into tools optimized for the manufacturing sector is our mission."