LISA | Integrating LLMs into existing software products

Region:

Flanders

Financed by

Practical guidance for secure, efficient and reliable implementation

Integrating large language models (LLMs) into existing digital products introduces new challenges in testing, security, operations, costs, and regulatory compliance. LISA helps Flemish software companies implement LLMs effectively and securely. The project provides practical methods and guidance to turn experiments into reliable software applications.

Target group

The project is primarily relevant for:

Software companies integrating LLMs into digital products and services
Digital product companies developing LLM-based functionality
Technology teams responsible for developing, testing, and operating LLM-based systems
Companies exploring LLM opportunities but lacking the internal expertise to implement them reliably and securely

Context

Generative AI and LLMs are changing how companies build digital products. Their ability to understand and generate natural language creates opportunities for customer interaction, content creation, data analysis, and personalised digital services. Integrating LLMs into existing software systems, however, introduces challenges that differ significantly from traditional software development.

LLM-based systems can behave unpredictably, making testing and regression management more complex. Companies must also manage hallucinations, changing models, operational costs, and security risks such as prompt manipulation and data leakage. Privacy, liability, copyright and compliance with the European AI Act add another layer of complexity. For many software companies, these challenges make it difficult to move from promising experiments to reliable production applications.

Objectives & results

LISA aims to help Flemish software companies integrate LLMs into digital products and services in a reliable, secure and sustainable way.

The project aims to:

Build practical knowledge on the implementation and integration of LLM-based applications
Develop methods for testing, deploying, monitoring and maintaining LLM-based systems
Identify security risks and provide guidance for secure-by-design implementation
Translate legal and regulatory requirements into practical guidance
Help companies evaluate technologies, costs, and dependencies when selecting LLM solutions

Key results:

Practical guidelines for implementing LLMs in software products, including architecture choices, tool evaluation and strategies to limit vendor lock-in
Security guidance addressing risks such as sensitive data disclosure, input manipulation, model theft, and system availability
DevOps and operational practices for testing, deployment, monitoring, reliability, and cost management
Legal and regulatory guidance covering privacy, liability, copyright, and the European AI Act
A lasting knowledge centre bringing together project insights, tools, and guidance for software companies

Approach

LISA combines the expertise of Sirris and DistriNet across four complementary research tracks:

LLM implementation
The partners investigate architectures, tools and frameworks for integrating LLMs into software applications. This includes technology monitoring and approaches to reduce dependency on individual providers.
Security
The project identifies security risks specific to LLM-based systems and investigates how secure-by-design principles can be incorporated from the start.
DevOps and operations
Existing software engineering practices are adapted to the specific characteristics of LLM-based systems, with a focus on testing, deployment, monitoring, reliability, and cost control.
Legal and regulatory aspects
The project translates requirements related to privacy, liability, copyright and the European AI Act into practical insights for software companies.

The knowledge developed across these four tracks will be brought together in a knowledge centre that remains accessible beyond the project duration.