Introducing Project Navigator: From AI intent to optimized deployment on Red Hat OpenShift AI

You've picked a model. Maybe it's a 70 billion parameter large model because someone on the team saw it top a leaderboard. Now you need it running in production on your Red Hat OpenShift AI cluster. So you start tuning batch sizes, figuring out quantization, sizing GPU requests, writing Kubernetes manifests, and hoping the out of memory errors stop before your deadline hits.

We've watched this play out enough times to see the pattern. The hard part of enterprise AI isn't just picking a model, it's the stretch between "this model looks good" and "this model is serving traffic reliably." That stretch eats weeks, sometimes months. Project Navigator is part of our strategy to make this process shorter.

The operational bottleneck

The AI industry has a model problem, but probably not the one you're thinking of. There are thousands of AI models available today. The tooling to train and fine-tune them keeps getting better. But getting a model deployed well, on hardware that fits, with a configuration that doesn't waste money? Typically, that still involves a lot of manual work.

GPU utilization at most organizations is well below what the hardware can do. Cost overruns on AI infrastructure are common. The people who know how to size and tune these systems are hard to find.

The pain shows up differently depending on your role. If you're an AI engineer, you spend weeks benchmarking models and debugging memory errors before a single request gets served. If you're a platform architect, you're watching expensive GPUs sit half-idle because nobody sized the workloads properly. And if you're the product owner, you can't tell whether a deployment will be cost effective until it's already running and the invoices start arriving.

What is Project Navigator?

Project Navigator is a layer on top of OpenShift AI that connects to your cluster, sees what's running and what's available, and helps you make better choices about model selection and deployment. You talk to it in plain language. Tell it what you're building ("I want a retrieval-augmented generation app for our internal knowledge base for 20 concurrent users with a max of 1.5s latency"), and it works with your cluster's actual state—the models in your catalog, the hardware you have, the benchmarks that matter for your use case.

Project Navigator is currently available in OpenShift AI 3.4 as a developer preview with 2 capabilities.

1. Intelligent model selection

Teams routinely pick models based on name recognition or leaderboard rankings without checking whether a smaller model can handle their task as or more effectively. Navigator changes this by matching what you describe against a collection of benchmark data (MMLU, HumanEval, and others) and recommending the best fit from your own model catalog.

The results can be surprising. For a code-heavy workload, an 8 billion parameter model may outperform a general purpose 70 billion parameter model on the benchmarks that actually matter, while using a fraction of the GPU resources. Navigator puts those comparisons in front of you so you pick based on evidence, not habit. It offers cost centric, performance centric, and blended recommendations.

2. Optimized model deployment

Once you've selected a model, Navigator generates Kubernetes manifests tailored to your cluster. That means a KServe InferenceService spec with resource requests and limits, a HorizontalPodAutoscaler with scaling rules, and a ServiceMonitor for Prometheus-based observability. All sized against what your cluster actually has, not generic defaults.

The gap between a properly configured deployment and a default one is large, both in hardware utilization and cost. Navigator closes it without requiring your team to become infrastructure specialists.

How it's built

Navigator is built on an open integration standard called the Model Context Protocol (MCP). This means it can meet teams where they work. If your engineers already use tools like Claude Code, Cursor, or Gemini CLI, they can access Navigator's capabilities directly from those environments. As more tools adopt MCP, Navigator will work with those as well, and as we bring Navigator's capabilities into Red Hat OpenShift AI, teams will have even more ways to access them.

The project is open source, developed upstream in the Open Data Hub community under the name RHOAI MCP Server.

Get involved

Project Navigator is in developer preview. We shipped it early because we want feedback from teams dealing with these problems right now.

Grab the code and try it: RHOAI MCP Server (Project Navigator)

If you're running OpenShift AI and spending too much time on model selection or deployment tuning, we'd like to hear from you. File issues, open pull requests, or tell us what's broken. We're building this in the open for a reason.