SatCom Expert Virtual Assistant (SCEVA)
An open science initiative by the European Space Agency’s ARTES programme, developed by Pi School in collaboration with RINA and i2CAT.
SCEVA is part of the SatcomLLM project (“Assessment of Open-Source Large Language Models within the SatCom Sector”), which aims to advance the use of Large Language Models (LLMs) for satellite communications.
SCEVA is a specialised Large Language Model system for satellite communications (SatCom).
It supports engineers, mission planners, and operators through natural language interaction, providing expert-level assistance in areas such as:
- Link budget analysis
- System design and simulation
- Constellation management and mission planning
- Cybersecurity and anomaly detection
- Regulatory compliance and documentation summarisation
The project combines scientific data curation, model training, benchmarking, and deployment to create open, reusable AI tools for the European SatCom community.
Two models are being fine-tuned using domain-specific data:
| Model | Description |
|---|---|
| esa-sceva/satcom-chat-8b | Compact SatCom-specialised model for inference and experimentation |
| esa-sceva/satcom-chat-70b | High-capacity model optimised for accuracy, reasoning, and domain coverage |
Training is carried out on EuroHPC and cloud GPU environments using Lit-GPT (Lightning AI), combining continued pretraining and instruction fine-tuning (IFT).
The models are being optimised for factual accuracy, reasoning depth, and integration with a Retrieval-Augmented Generation (RAG) pipeline.
All datasets are curated and structured for reproducible research and fine-tuning.
| Dataset | Description |
|---|---|
| esa-sceva/satcom-qa | Expert question–answer pairs on SatCom workflows |
| esa-sceva/satcom-mcqa | Multiple-choice questions for evaluation and benchmarking |
| esa-sceva/satcom-synth-qa | Synthetic QA dataset generated using SCEVA’s agentic data pipeline |
| esa-sceva/satcom-synth-qa-cot | Chain-of-thought reasoning dataset for improved interpretability |
Datasets include contributions from open-access technical sources, ESA documentation, and curated domain corpora compliant with EU and ESA data guidelines.
SCEVA integrates:
- LLM Core: Fine-tuned LLaMA-based models (8B and 70B)
- RAG System: Hybrid document retrieval with vector databases for source-grounded answers
- Agentic Layer: Task automation and tool-calling via LangGraph
- Web Interface: Internal demo interface and API for expert users
All released models, datasets, and tools are open access and follow ESA and EU open science guidelines.
The goal is to enable European technological sovereignty in AI for SatCom and to foster a shared ecosystem for applied space AI research.
Contributors:
Pi School · RINA Consulting · European Space Agency (ARTES Programme)
