Date: 25 September 2026
Time: 10:30–12:00 CEST
Participation: Open to all ESA Member States. Registration is required by 25 September at 00:00 CEST. Registration now open.

---

INTRODUCTION

How can you optimise an open-source Large Language Model (LLM) for satellite communications applications and customise it using an organisation’s own technical information, while maintaining confidentiality?

This practical webinar will showcase the results of the SatComLLM activity and demonstrate the complete workflow used to develop SCEVA-RAG, the SatCom Expert Virtual Assistant combining: Fine-tuned SatCom Llama + Retrieval-Augmented Generation (RAG) = SCEVA-RAG.

Participants will see how a general-purpose Llama model can be adapted to the satellite communication domain using Low-Rank Adaptation (LoRA) fine-tuning, and how Retrieval-Augmented Generation (RAG) can connect the model to a searchable knowledge base built from technical documents.

The session will follow the complete development journey, from collecting and preparing satellite communication literature to generating training data, fine-tuning the model, creating a document knowledge base and integrating the components into an operational assistant.

WHAT WILL BE DEMONSTRATED?

The webinar will include practical demonstrations of:

• Building a high-quality satellite communication corpus from technical literature
• Converting documents into training data and searchable knowledge
• Generating and quality-checking synthetic satellite communication question-and-answer examples
• Configuring the two-stage LoRA fine-tuning workflow
• Creating a knowledge base from technical PDF documents
• Connecting a fine-tuned satellite communication model to a RAG pipeline
• Comparing model responses with and without document retrieval
• Deploying the solution locally, in the cloud or through a hybrid architecture

A complete production fine-tuning run of an 8-billion or 70-billion parameter model requires several hours and substantial computing resources. The webinar will therefore demonstrate the fine-tuning configuration, execution process, outputs and resulting model artefacts, using completed satellite communication model checkpoints for the live SCEVA-RAG demonstration.

AGENDA

---

1. INTRODUCTION AND SYSTEM OVERVIEW | 10 MINUTES

• Overview of the final SatComLLM architecture
• Main steps followed to develop the satellite communication-specialised models
• Relationship between fine-tuning, the knowledge base and RAG
• Introduction to the complete SCEVA-RAG workflow

---

2. CREATING THE SATELLITE COMMUNICATION CORPUS, TRAINING DATA AND EVALUATION DATASETS | 10 MINUTES

• Collection and preparation of satellite communication literature
• Document extraction, cleaning and chunking
• Generation of synthetic question-and-answer data
• LLM-based filtering and quality assurance
• Introduction to the satellite communication-QA and satellite communication-MCQA evaluation datasets, with practical examples

---

3. FINE-TUNING WORKFLOW | 15 MINUTES

• Practical walkthrough based on the delivered fine-tuning tutorial
• Selection of the Llama base models
• Two-stage LoRA fine-tuning methodology
• Training configuration and execution
• Review of training outputs and the resulting SatCom-8B and SatCom-70B models
• Infrastructure requirements and indicative training runtimes

---

4. KNOWLEDGE-BASE CONSTRUCTION AND RAG INTEGRATION WITH PRIVATE DOCUMENTS | 15 MINUTES

• Practical walkthrough based on the delivered RAG tutorial
• Loading a collection of technical PDF documents
• Document chunking and embedding generation
• Creation of a searchable knowledge base
• Explanation of document retrieval and contextual grounding
• End-to-end SCEVA-RAG question-and-answer demonstration
• Comparison of responses with and without retrieval augmentation

---

5. DEPLOYMENT CONSIDERATIONS | 5 MINUTES

• Local, cloud and hybrid deployment options
• Privacy, data sovereignty and infrastructure trade-offs
• Hardware considerations for the SatCom-8B and SatCom-70B models
• Deployment considerations for proprietary or sensitive technical information

---

6. QUESTIONS AND DISCUSSION | 20 MINUTES

Participants will have the opportunity to discuss the technical approach, potential use cases and options for adapting the SatComLLM ecosystem to their own data and operational environments.

---

WHO SHOULD ATTEND?

The webinar is intended for:

• Companies and institutions from all ESA Member States
• Satellite communications engineers and system architects
• Artificial intelligence and machine-learning practitioners
• Technical and knowledge-management teams
• SatCom operators, manufacturers and service providers
• Research organisations and universities
• Organisations considering secure, domain-specific AI assistants
• Teams interested in fine-tuning open-source models or implementing RAG with proprietary technical documentation

WHAT WILL PARTICPANTS LEARN?

By the end of the session, participants will understand:

• The difference between fine-tuning and RAG, and why the two approaches are complementary
• How satellite communication documents can support both model training and a RAG knowledge base
• How two-stage LoRA fine-tuning adapts a general-purpose Llama model to satellite communication
• How RAG provides documentary grounding, traceability and access to updated or private information
• How the fine-tuned model and retrieval pipeline are integrated into SCEVA-RAG
• The practical infrastructure and deployment choices involved in implementing a similar solution

Join us for a practical demonstration of how open-source LLMs, domain-specific fine-tuning and document retrieval can be combined to create a trusted satellite communication AI assistant.