Building a private LLM for satellite communications

​Satellite communications engineering depends on technical standards, link budgets and calculations where precision matters. General-purpose language models, especially smaller ones, often get these wrong: they produce incorrect formulas or invent values. For anyone designing or analysing satellite systems, that is a reliability problem.

​SatcomLLM was built to solve it. Developed by Pi School and RINA Consulting within the European Space Agency’s ARTES programme, it is a private assistant made for satcom. It combines two open models, a public corpus, human-written benchmarks and a RAG system that answers from your own documents and cites them.

​In this 75-minute online session, deep learning scientists Vijayasri Iyer and Riccardo Corrente will show how they built it, from the data to deployment. They will explain how the team turned 169,527 open-licence documents into a corpus that can be searched and released, and how they generated and filtered 182,756 question-and-answer pairs for training. They will also cover how 1,990 human-written questions were kept aside to test the model fairly. In live notebooks, they will walk through two-stage LoRA fine-tuning on Llama 3.1 8B and Llama 3.3 70B, then show the same question answered with and without RAG on private documents. The session closes with the trade-offs between local, cloud and hybrid deployment, followed by 20 minutes of Q&A.

​The lessons apply to any organisation whose knowledge is too specialised for general-purpose AI, or too sensitive to leave its own infrastructure. Everything shown is public on GitHub and Hugging Face, so you can start building the same day.