$100k+
Azure and AI credits
Azure
Production deployment
Foundry
Models in the stack
What the programme gave
Azure and AI credits worth more than $100k, which is the difference between running an experiment once and running it until the result means something.
With it came access to Microsoft's AI services and developer tooling, and the documentation, guides and support behind the replication and modernisation paths.
Engineering
Caracal on Azure
- Problem
- Security infrastructure has to run somewhere companies can actually adopt it, not just on my machine.
- Approach
- Deployed Caracal end to end on Azure and served it from there, with models from Azure AI Foundry behind the agent-facing parts of the system.
- Result
- A deployment other companies could be served from, and enough headroom to run adversarial evaluation repeatedly rather than sparingly.
AzureAzure AI FoundryDeploymentModel evaluation
Evaluation at scale
You cannot claim a policy engine is sound because it worked locally. It has to survive adversarial workloads, repeatedly, and that costs compute.
The part I used hardest
The guidance and technical support, not the credits. Engineers who have shipped platform software at a scale I have not, willing to say where the design would break.
Two architectural decisions exist in their current form because someone there pushed back on my first answer.