HelixML, a start-up based in Bristol, UK, introduced a local Generative AI (GenAI) solution aimed at enterprises prioritizing data sovereignty and regulatory compliance. The platform allows organizations to use cutting-edge AI technologies while ensuring sensitive data remains within their own infrastructure, addressing concerns around data privacy and external AI providers like OpenAI. HelixML’s architecture leverages open-source models such as Meta’s Llama3.1, paired with a flexible GPU scheduler, enabling efficient deployment and management of multiple large language models (LLMs).
HelixML’s design focuses on providing enterprises with a secure, customizable, and scalable AI solution. Its fine-tuning capabilities allow organizations to tailor models to their unique needs without the risk of data leaving their premises. The system also supports Retrieval-Augmented Generation (RAG) to integrate external knowledge sources.
The launch of Helix 1.0 follows nine months of development. CEO Luke Marsden highlighted the importance of adopting AI securely, stating, “Every major company is looking at GenAI, but they need to adopt it in a way that doesn’t compromise their data security.”
- Data sovereignty and compliance: HelixML ensures that enterprises can utilize AI without data leaving their own infrastructure, addressing the needs of highly regulated industries like finance and healthcare.
- Open-source AI models: HelixML leverages open-source large language models such as Meta’s Llama3.1, enabling companies to benefit from advanced AI technologies while maintaining full control over their data.
- High-performance GPU scheduler: The platform’s architecture includes a GPU scheduler that allows enterprises to mix and match different AI models, optimizing hardware resources and ensuring efficient model deployment and operation.
- OpenAI-compatible API: The system integrates easily with existing applications via an OpenAI-compatible API, allowing businesses to deploy AI models without significant changes to their current infrastructure.
- Security and privacy: HelixML’s design ensures security with in-VPC (virtual private cloud) deployments and features that prevent data leakage, making it suitable for enterprises requiring strict data privacy and security.
- Control Plane architecture: The platform’s control plane, built using Golang, serves as an API gateway and reverse proxy, managing interactions between components. It communicates with runners (which handle GPU tasks) via API and WebSocket connections.
- Runner system: HelixML’s runners manage GPU resources by polling the API server for jobs and executing them. These containerized runners ensure efficient use of GPU memory and can run securely behind network address translation (NAT).
- RAG and knowledge integration: The system’s RAG capabilities enhance AI outputs by incorporating external data, making it useful for enterprises that need to include specific knowledge bases in their AI-driven operations..