This case study covers a trusted IT security provider for critical government services. They are the go-to provider for delivering AI/ML projects for security conscious public sector agencies and private sector companies. Their customers have come to expect high-velocity innovation and development without compromising on security.

As a leader in secure IT services, their goal was to standardize and secure every stage of their AI/ML development lifecycle — from experimentation to deployment — while ensuring compliance, reproducibility, and operational scalability.

With this in-house development team, they considered building their own tools, or extending existing ones, but chose instead to adopt KitOps ModelKits for packaging, versioning, and storing both their customers’ and their own internal AI/ML projects. This immediately saved them weeks of time by being able to use an off-the-shelf open source project rather than build their own tooling. However, the larger savings came with each of their project cycles. Since using KitOps they have seen project times reduced by over 13% per cycle.

As a next step, adding the Jozu Hub will accelerate development cycles by an additional 34%.

A Lack of Central Control

To cover everything from data preparation, to model training, to testing and deployment to inference servers, the customer had brought together a dozen tools including popular tools like Jupyter Notebooks, MLFlow, and DvC. They also had existing tools used for application management like git, CI/CD pipelines, and Kubernetes.

Their portfolio of AI/ML tooling, while functional, didn’t provide the security and compliance their customers demanded. As is common in AI/ML, their toolchain was made up of a half-dozen tools which didn’t tightly integrate with each other - adding delays and risk to handoffs and coordination within their teams.

Their initial AI/ML projects had been successful, but lacked the centralized control and compliance that the high security organizations they worked with demanded:

• Fragmented Asset Management Sensitive assets like datasets, models, and code were stored across disparate systems (Git, Amazon S3, Jupyter Notebooks, MLFlow, and Kubernetes).
• Security Gaps Existing tools and storage solutions were vulnerable to tampering, lacked fine-grained and consistent authorization, and had no unified versioning or auditability.
• Sluggish Handoffs Development and deployment teams struggled to move models from experimentation to production efficiently and securely.

These limitations caused delays, created needless compliance risks, and hindered the customer’s ability to scale their secure AI/ML offerings.

They considered forking the open source Jupyter and MLFlow projects to add secure storage to each independently but quickly dismissed the idea because of the high initial and on-going investment it would require from them to build and maintain these bespoke tools.

How They Will Further Improve their Security and Speed with Jozu Hub

As a next stage, the customer is planning to use the Jozu Hub as their secure service catalogue for AI/ML projects. This will allow them to easily stay on top of public open source model advances, while ensuring that their development teams are only using base models that have acceptable licenses, lineage, and security posture.

Storing KitOps ModelKits in Jozu Hub will give them even greater benefits:

• A centralized service catalogue for AI/ML projects, surfacing approved datasets and models with metadata and lineage.
• Guided paths for development teams making it easier for teams to use approved and vetted ModelKit-packaged datasets and models than pull them from Hugging Face where provenance and security are unknown.
• Vulnerability scanning and provenance tracking for every model and pipeline stage.
• Automated model deployment generation reducing handoff time and ensuring consistency across Kubernetes environments.

Together, KitOps and Jozu Hub will give the customer complete visibility, control, and agility over their AI/ML lifecycle — from data ingestion to production inference.