Accelerating Syncable’s specialised AI models for automated cloud deployment

The organisation

Syncable is a Danish AI startup, backed by Innovation Fund Denmark, specialised in bridging the gap between code generation and multi-cloud deployment, including EU sovereign cloud vendors. Syncable combines expertise in AI, commercial strategy, and full-stack development to build a platform where AI agents analyse code repositories to automatically generate infrastructure-as-code assets like Dockerfiles and Terraform scripts and automate the cloud deployment process with production grade quality. Their solution enables seamless deployment to any environment, cloud, on-premise, or HPC networks, utilising specialised small language models to ensure accuracy and reduce dependency on general commercial LLMs.

The challenge

General-purpose large language models (LLMs) are often overkill for specific infrastructure tasks, resulting in slow performance, high costs, and a significant CO₂ footprint. Generating critical deployment files such as Terraform configurations and Dockerfiles requires strict adherence to best practices and zero tolerance for hallucinations. Syncable also needed to address the challenge of verifying AI-generated code to ensure security and reliability in deployment, a task that requires deep knowledge extraction via knowledge graphs. Developing and training dedicated small language models (SLMs) is resource-intensive and prohibitively expensive on commercial cloud providers.

The solution

The project utilised EuroHPC GPU resources on LUMI to train dedicated small language models specifically for generating infrastructure code (Terraform, Docker), optimised for speed and accuracy rather than general conversational ability. The solution also involved developing AI agents capable of deep code analysis using knowledge graphs to map code relationships and reduce hallucinations.

Business impact

Training specialised models gave Syncable a competitive advantage, enabling the company to offer a faster, cheaper, and more accurate deployment solution compared to competitors relying on generic LLM APIs. The approach also delivers a significantly lower CO₂ footprint, making it more environmentally sustainable.

Benefits

• Cost efficiency: Specialised SLMs reduce computational overhead and cost compared to large general models
• Reliability: Verifiable AI agents reduce deployment errors and security risks
• Scalability: Validated HPC workflows allow the startup to rapidly scale training as new deployment targets are added

All use cases