Case studies

Why Satellite Communications Needs Its Own Custom LLM

Satellite communications is a domain where precision matters. Engineers working on link budgets, mission documentation, and anomaly analysis need answers that are accurate, verifiable, and grounded in trusted technical sources. In highly specialised fields like this, a general-purpose model is simply not enough. Building custom LLMs through domain-specific fine-tuning is what makes the difference.

This is precisely where Pi School’s expertise comes in. In less than a month, Pi School has been involved in the release of two open-source custom LLMs built for high-stakes applications with the European Space Agency.

One is EVE, the first Earth Virtual Expert developed for ESA Φ-lab in partnership with Imperative Space, Mistral AI, and Wiley. EVE brings Earth observation and Earth science knowledge to anyone who needs it.

And now SatcomLLM, developed under ESA’s ARTES programme in partnership with RINA, brings the same approach to satellite communications. Building vertical custom LLMs that are open, reliable, and deployable in sensitive environments is one of the AI research fields where Pi School excels. These two projects are proof of that.

Why Satellite Communications Needs Its Own Custom LLM

What makes SatCom Expert Virtual Assistant – SCEVA – different

SCEVA is a custom LLM built on an open-source foundation and fine-tuned on a curated corpus of approximately 170,000 satellite communications documents, covering everything from technical standards to mission workflows. Unlike proprietary systems, custom LLMs like SCEVA can be deployed locally, on the cloud, on local servers, or on edge platforms, giving organisations full control over sensitive data. This makes it particularly relevant for institutional and industrial stakeholders operating in regulated environments.

To address the challenge of factual reliability, SCEVA integrates retrieval-augmented generation, grounding every response in trusted documents from a dedicated knowledge base. Users can also populate this knowledge base with their own proprietary materials, enabling further custom adaptations without compromising data security.

Two model variants are available, an 8B and a 70B parameter model, both fine-tuned in two stages. The first stage covers broad satcom terminology and workflows. The second focuses on reasoning and multi-step problem solving, training the custom LLM to plan, justify, and compute.

Why custom LLMs matter for regulated industries

The models, datasets, benchmarks, and codebase are all released under open licences, contributing to European digital sovereignty and reducing dependence on non-European AI providers. The entire pipeline is publicly available on GitHub and Hugging Face. For organisations working at the intersection of space technology and AI, SatcomLLM demonstrates what purpose-built custom LLMs can deliver in practice: domain depth, factual reliability, and full data control.

Explore the project: