A leading research university's faculty increasingly rely on large language models for data analysis, software development and scientific computing. Researchers wanted to experiment with both open-source and commercial models, without configuring GPU drivers and AI runtimes themselves.
The university also wanted commercial models consumed through Amazon Bedrock rather than external services, with institutional control over which models were available and clear visibility into what AI usage was costing.
Relevance Lab deployed AI Workspaces within Research Gateway, giving researchers one environment that runs open-source models locally and calls Amazon Bedrock when they need it.
Case study • PDF
AI Workspace Architecture Blueprint
Researchers wanted to test open-source and commercial models without configuring GPU infrastructure themselves. See how AI Workspaces put Ollama and Amazon Bedrock in one governed environment, with costs visible per project.
View the Full Case Study →
Business Challenges
The university wanted to democratize access to AI technologies while simplifying infrastructure management and maintaining institutional governance.
01
Supporting Multiple AI Model Types
Some projects called for open-source models running locally on GPU infrastructure through Ollama. Others needed approved commercial models through Amazon Bedrock. The university wanted one platform that did both.
02
Simplifying AI Infrastructure
Setting up GPU drivers, AI frameworks, model runtimes and development tools took specialized expertise. Research IT wanted researchers productive on day one.
03
Flexible GPU Computing
Researchers needed GPU infrastructure only when a workload called for it, and the ability to pick the right configuration each time, so performance did not come at the cost of the cloud bill.
04
Secure Enterprise AI Access
Commercial models had to be reached through Amazon Bedrock rather than external services, with Research IT deciding which foundation models were available and every API call staying inside the university's AWS environment.
05
AI Cost Visibility
Administrators needed Amazon Bedrock API consumption reported alongside infrastructure costs, attributed to the research workspace that generated it.
Solution
Relevance Lab deployed Research Gateway with a new AI Workspaces catalog offering that enabled researchers to provision fully configured AI development environments through a self-service portal.
Researchers launched AI-ready workstations on demand, fully configured on arrival.
- Browser-based remote desktop
- CPU and GPU-enabled Amazon EC2 instances
- NVIDIA GPU drivers pre-installed
- Standardized software images
- Persistent research storage
- Self-service lifecycle management
Every workspace ran Ollama as a Docker container for open-source LLMs held locally.
- Local API endpoint
- Inference stays inside the workspace
- Multiple open-source models supported
- Model chosen at provisioning, installed automatically
- No external AI dependencies
A standardized research toolset, pre-installed and working on first login.
- Python and R
- JupyterLab and RStudio
- Git
- Ollama runtime
- OpenWebUI, a browser chat interface for local models
Each workspace reached Amazon Bedrock through IAM roles scoped to approved models.
- IAM-based model access
- Foundation models approved by Research IT
- Claude and other Bedrock-hosted models
- API requests stay inside the university's AWS environment
Research Gateway mounted approved project datasets into every AI Workspace automatically.
- Study folders mounted at launch
- Existing project permissions applied
- Project isolation maintained
- Governed access to approved datasets
Financial reporting extended to cover Bedrock consumption alongside infrastructure spend.
- Infrastructure and GPU compute costs
- Amazon Bedrock API costs
- Workspace operating costs
- Project-level AI spending
Architecture Highlights: Unified AI Research Platform
Researchers authenticate to Research Gateway and provision workspaces from a catalog Research IT controls. Each workspace sits in its own governed project environment, with two inference paths available from the same desktop.
Access through Research Gateway, with catalog offerings approved by Research IT
Workspaces provisioned into isolated project environments, study folders mounted at launch
Local inference through Ollama, running entirely on the workspace GPU
Amazon Bedrock inference through instance IAM roles, staying inside the university's AWS account
Usage from both inference paths reported together, per workspace and per project
Business Outcomes
Researcher Productivity
- GPU-enabled workspaces provisioned on demand
- Several foundation models testable from one environment
- No manual driver or runtime setup
Flexible AI Innovation
- Open-source models running locally through Ollama
- Approved Bedrock models from the same workspace
- OpenWebUI chat alongside JupyterLab and RStudio
Operational Efficiency
- Self-service provisioning from a governed catalog
- Standardized images maintained centrally
- Bedrock model access controlled in one place
Financial Visibility
- Bedrock API usage tracked per workspace
- AI costs allocated to research projects
- GPU utilization visible alongside AI spend
Customer Perspective
“
Research Gateway let us deliver enterprise-grade AI capabilities to our researchers without introducing unnecessary operational complexity. By combining open-source models, Amazon Bedrock, GPU computing, and self-service provisioning into a single platform, we have accelerated AI adoption while maintaining governance, cost visibility, and institutional security.
— University Research Computing Leadership
Research Gateway for AI-Enabled Research
Research Gateway gives researchers governed access to AI without forcing a choice of vendor. Open-source models run locally on the workspace GPU. Approved commercial models come through Amazon Bedrock. Both report costs into the same project view.
AWS recognized Research Gateway as one of the top two partner solutions globally for Higher Education. Universities come to us to give researchers self-service AI environments without Research IT rebuilding the stack for every project.
AI-Enabled Research Case Study
Want Researchers Choosing Models, Not Configuring GPUs?
See how one research university put open-source models and Amazon Bedrock in the same GPU workspace, with usage from both reported against the project that ran it.
