From idea to orbit: Terra Mater Studios builds Hera AI companion with Azure data, AI

In the expanse of space, a new kind of intelligence is emerging—and a new model for human-AI collaboration. It started with a bold idea from a creative visionary. “What if we could speak with a spacecraft like it was a friend?” says Markus Mooslechner, Concept and Creative Lead at Terra Mater Studios, a nature film and documentary production company and subsidiary of Red Bull. “Not through cold commands or scripted chat, but through meaningful, evolving conversation.”

A longtime science communicator and creator of documentaries, Mooslechner wanted people on Earth to have a personal connection with space. In 2023, his vision found a home in the Hera spacecraft mission of the European Space Agency (ESA). It was set to embark for the Didymos binary asteroid system the following year to gather data for a planetary asteroid defense initiative. “It was ideal—an exploratory mission that people could speak to,” says Mooslechner.

Ian Carnelli, Hera Mission Manager at ESA, loved the idea, but knew building a secure, natural conversational AI for a spacecraft traveling 6.1 kilometers per second to a destination 195 million miles away wouldn’t be easy. “Nothing like this has ever been done. It’s like science fiction, but we knew Microsoft AI technology could make it real,” says Carnelli.

A foundation of safety

Terra Mater Studios, ESA, Microsoft, and Impact AI teamed up to develop a public-facing AI platform linked to a spacecraft, where safety would be paramount. “ESA’s number one concern was: don’t let this thing go rogue,” says Mooslechner. “We had to ensure it wouldn’t become another HAL 9000 from the film ‘2001: A Space Odyssey.’”

The team chose to build on the highly secure foundation of Microsoft Azure AI Foundry. “Using Azure AI Foundry was essential, as it gave us a governance backbone to operationalize quickly and deploy models safely at scale. It’s not just infrastructure—it helped accelerate a safety-first approach across the full development life cycle,” says Anna Maria Brunnhofer-Pedemonte, CEO of Impact AI. “Together with our Impact AI platform, the entire stack is designed for testing, evaluation, and alignment of complex compound AI systems like Hera.”

Relying on a layered evaluation and safety stack, the team built trust into every interaction. Hera’s responses were scored in real time across factuality, tone, and relevance, and identified when to regenerate or block content. In addition, custom agents in the Impact AI platform track engagement, sentiment, coherence, and brand alignment, while Azure AI content safety systems—like Prompt Shields—filter every input and output. Offensive prompts or unsafe requests never make it to model inference. “We score every message on factuality, brand tone, robustness, and alignment with Hera’s personality,” says Justin Bercich, Chief Technology and Product Officer at Impact AI.

The platform also provides a helpful visual layer. Tracing, annotation, and prompt debugging tools helped the team optimize conversations and fine-tune model behavior based on live telemetry from ESA or dynamic public input.

Azure AI Foundry also supports fast iteration. “When ESA updates telemetry endpoints or Azure OpenAI releases new models, we can redeploy within hours,” says Brunnhofer-Pedemonte. “That kind of agility is imperative in live mission environments.”

Engineering a conversational system from scratch

In early 2024, building a production-grade conversational agent meant charting unfamiliar territory. “There were no stable frameworks for orchestrating agents, managing persistent memory, or real-time guardrails,” recalls Bercich. “This isn’t a dashboard. It’s a live, reasoning agent in space—we had to build it from scratch.”

The team started with Visual Studio Code to write and manage code for the foundational architecture. Azure Container Apps hosted three microservices: the UI front end, an orchestration back end, and an external API bridge—all deployed using serverless consumption plans to facilitate elastic scaling. Azure API Management served as both external gatekeeper and internal router, splitting traffic to help ensure both security and performance while directing requests to specific AI tools, like Azure AI Search, or to Azure Cosmos DB.

“Azure Cosmos DB became Hera’s memory engine,” says Brunnhofer-Pedemonte, “managing session state, user preferences, and summarizing conversations.” Azure AI Search within Azure AI Foundry powered semantic retrieval-augmented generation (RAG), synthesizing insights from ESA’s mission archives. “We ingested hundreds of ESA documents, videos, and telemetry,” Brunnhofer-Pedemonte says. “Then we tagged and chunked them for efficient RAG. Real-time latency was essential—and Azure AI Search delivered.”

Azure OpenAI in Azure AI Foundry Models brought the models to life, offering low-latency endpoints and built-in content filtering through Azure AI Content Safety.

Each query can involve up to 10 language model calls. Hera plans, reflects, and decides whether to retrieve knowledge, recall memory, or ask a follow-up before responding. “That orchestration logic is purpose-built,” says Bercich. Hera’s planning agent determines which tools to activate, while a reflection agent assesses tone, flow, and relevance. Context is stored and retrieved via summarized context windows in Azure Cosmos DB, preserving coherence across fragmented sessions.

Securing the pipeline, end to end

Security was critical. “Every layer, from ingestion to delivery, had to be fortified,” notes Carnelli. API Management separated internal and public services, isolating telemetry APIs. Bicep provisioning and GitHub Actions enforced governance across multiple tenant environments, spanning test, staging, and production. “We optimized developer workflows with CI/CD pipelines using GitHub Actions and scaled applications with Azure Container Apps,” says Brunnhofer-Pedemonte.

The Azure Cosmos DB serverless model helped ensure high availability with SLA-backed uptime, while its partitioning logic enabled precise control over sessions and anonymized historical data. Storage stayed compliant with ESA’s data sovereignty policies, with deployment anchored in Sweden Central, a Microsoft cloud region.

“We evaluated other databases, but Azure Cosmos DB gave us the scale, speed, and interoperability we needed—without extra overhead,” explains Bercich.

The following diagram shows the architecture of Terra Mater’s Hera AI companion app built on Microsoft Azure. It outlines the layered structure of the app, including data sources, storage, governance, processing, and user interface components.

Speaking with a spacecraft

The Hera AI companion was created in a few months, and the spacecraft launched in October 2024. The public Hera.Space portal makes it possible for anyone to chat with Hera. She speaks multiple languages, adjusts tone for different audiences, and even has a sense of humor.

“We’re syncing live with ESA telemetry, and she translates it in real time,” says Bercich. “You can ask Hera what she sees, how the solar panels are functioning, or what happened when her orbital thruster glitched—and she answers.” She even narrated her own Mars flyby.

Hera also offers unsolicited insights. Carnelli recalls an ESA briefing where, out of the blue, the AI specifically gave a shout-out to Brian May, the guitarist from Queen, who is also an astrophysicist and contributor to the mission. The AI made an unscripted recognition based on prior knowledge and context. “That was a turning point” says Mooslechner. “It reminded us that intelligence—even artificial—can surprise you.”

A new way of thinking and building

As space missions grow more complex, Hera signals a new era: a companion that learns, adapts, and reflects—not just a chatbot, but “a new voice in the solar system,” says Carnelli.

“We’re thrilled to be on this journey with Microsoft,” adds Brunnhofer-Pedemonte. “The possibilities are limitless. The key is to keep pushing. With Azure, we’re at the cutting edge.”

Bercich adds, “If you’re a developer, use Azure AI technology to build fast, test rigorously, and use evaluations as your North Star. That’s how you scale trust.”

“And never forget the why,” Mooslechner concludes. “Technology alone doesn’t move people. A story does. The Hera companion is undoubted proof that when expert AI solutions and engaging storytelling come together, magic happens.”

Discover more about Terra Mater Studios on Instagram, LinkedIn, Facebook, and YouTube.