ELAN Languages Scales AI-Powered Translation with zwrm Sandboxes

The Challenge

ELAN Languages provides AI-assisted translation and language learning tools to companies and institutions across Europe. Their LLM-powered translation agents regularly need to process sensitive content — but running arbitrary model inference in shared container environments raised serious security and compliance questions.

"Our customers are governments, police stations and regulated businesses. They need to know that client data is processed in complete isolation," explains ELAN Languages CIO, Tom Jordi Ruesch.

Key pain points

• Container-based isolation insufficient for processing sensitive personal data
• Scaling LLM inference workloads was expensive on traditional cloud providers
• GDPR compliance requirements for client data across multiple EU member states
• Need for fast iteration on agent prompts and model configurations

The Solution

ELAN Languages deployed their translation agent fleet on zwrm, using Firecracker microVM sandboxes to achieve hardware-level isolation for each agent execution. Each translation job runs in its own ephemeral VM — booting in milliseconds, processing the content, and shutting down cleanly.

Architecture

The system runs a pool of zwrm machines that receive translation jobs from a central queue. Each job spins up a sandboxed agent that:

1. Loads the appropriate language model configuration
2. Processes the source text through the translation pipeline
3. Returns results and terminates — no state persists between jobs

# zwrm.toml for the translation agent worker
[app]
name = "elan-translate-worker"

[vm]
size = "performance-2x"

[[services]]
internal_port = 8080
protocol = "tcp"

The Results

After deploying on zwrm, ELAN Languages saw significant improvements in both performance and cost efficiency.

• 3x throughput improvement — Firecracker's lightweight VMs boot faster than containers, enabling rapid job cycling
• Hardware-level isolation — Every translation job runs in its own microVM, satisfying even the strictest data protection requirements
• 40% infrastructure cost reduction — Efficient resource utilisation compared to their previous dedicated GPU instances
• 24 language pairs — Scaled from 12 to 24 supported language pairs without infrastructure changes

"With zwrm, we can tell our public and private sector clients that every single translation job runs in complete hardware isolation. That's not marketing — it's the actual architecture."

— Tom Jordi Ruesch, ELAN Languages

What's Next

ELAN Languages is shifting much of its internal agent-led development process to the zwrm ecosystem, using zwrm's agents, machines and low-latency sandboxes to centralize everything from experimentation, to website hosting to custom AI deployments on a single platform.