Air Cloud Customer Case Study: An NPU-Powered Private AI Platform Built in Two Weeks
“We want to add AI to our services. But where should we process customer data and internal documents—and can we use the hardware we already own?”
For enterprise teams, choosing an AI model is only part of the decision. The environment it runs in matters just as much. Public cloud can help teams provision resources quickly and scale with demand. Private cloud offers direct control over data and operations, while a hybrid approach connects the two. When security and data requirements lead a company to run AI on its own infrastructure, another challenge follows: making models accessible to the teams and applications that need them.
AIEEV built an NPU-powered private AI platform for Company A, Korea’s largest AI contact center (AICC) platform provider, in just two weeks. Built on Private Air Cloud, the deployment combined model serving on neural processing units (NPUs) with large language model (LLM) routing. This gave internal teams a way to call and use multiple models. Here is how we connected the company’s NPU infrastructure to an AI platform its internal applications could use.
1. Connecting model execution to internal applications
When testing a language model, teams usually focus on output quality: Does it understand the question? Can it summarize a long document? Does it follow the requested format? Integrating that model into an internal application adds another question.
“How does our application send requests to this model?”
Consider a document summarization feature. Beyond manually testing a model, the application needs to send a document, receive a summary, and display it to the user. That requires a model-serving environment that runs the model and handles requests from other applications. On NPU infrastructure, teams also need to check model execution requirements, since supported models and software configurations can vary by hardware. For Company A, we treated model execution and internal access as one deployment scope: running models on NPUs and connecting them to the applications that would call them.
2. Deploying Private Air Cloud on the customer’s NPU infrastructure
Private Air Cloud is AIEEV’s solution for building a private AI cloud on an enterprise’s own GPU and NPU infrastructure. It provides a platform for deploying and using AI models within the customer’s environment.
We used Private Air Cloud as the foundation for Company A’s NPU-powered private platform, then configured model serving so internal applications could send requests and receive responses.
Using an existing platform as the foundation allowed us to focus on the customer’s model execution environment and integration needs. Alongside NPU model serving, we configured LLM routing to connect requests to the models the applications intended to call.
3. Routing internal requests to multiple models
With one model, the request destination is straightforward. As teams begin evaluating or using multiple models, they need a way to address each one from their applications. For example, comparing summaries of the same document requires sending requests to different models and receiving their responses. We configured an LLM routing platform for Company A to forward internal requests to the requested model and connect them to the model-serving environment. The request flow is:
Internal application request → LLM routing → Model serving → Response
Serving and routing perform distinct roles. Model serving runs the model and processes the request; routing forwards the request to its destination. Connecting the two gave Company A’s internal teams access to multiple models. Developers could retrieve responses and evaluate how to incorporate them into internal tools, including features such as document summarization.
4. A private AI platform, model serving, and LLM routing—in two weeks
Private AI deployments involve several connected tasks. Depending on the project, the scope may include model deployment, access management, integration with existing systems, and user testing. For example, GreenM’s private AI search assistant for NPH Group was delivered over six weeks, including LLM and vector storage deployment, integration with an existing portal, user testing, and training. These projects illustrate the work involved in connecting models to their users. Their scopes and environments differ, so timelines alone are not a like-for-like speed comparison.
For Company A, we configured the Private Air Cloud platform, NPU model serving, and LLM routing together in two weeks.
NPU-powered private platform: Deploying Private Air Cloud in Company A’s NPU environment.
Model serving: Configuring the environment to run models and process requests.
LLM routing: Connecting internal requests to the models they are intended to call.
Together, these three components enabled Company A’s internal teams to call and use multiple models. Private Air Cloud supplied the platform foundation, while customer-specific model serving and routing connected the NPU infrastructure to internal applications.
Building private AI around your infrastructure
Every enterprise has a different mix of models, compute resources, and internal applications. Planning a private AI environment means considering all three together. In this project, AIEEV connected those components in two weeks, building an NPU-powered platform with model serving and LLM routing for Company A.
Planning to deploy AI models on your own GPU or NPU infrastructure, or connect them to internal services? Talk to AIEEV about a private AI environment built around your hardware and use cases.
.
.
.
