NEXAARIS LABS / RADAR
Technology radar: our view of the Microsoft and AI ecosystem.
Our opinionated assessment of the platforms, tools and techniques we recommend, trial, watch or avoid — based on hands-on use in Nexaaris Labs and client delivery. Reviewed periodically.
How to read the radar
Adopt means we recommend it as a default for appropriate workloads. Trial means it is worth using on a project that can manage some risk. Assess means worth exploring to understand its effect. Hold means proceed with caution or avoid for new work. Positions reflect our experience and will change as technology matures.
Adopt
Our defaults for new work on the Microsoft platform.
- .NET 10 (long-term support) for new services and applications
- Azure Container Apps for containerised services that do not need direct Kubernetes control
- Microsoft Entra managed identities and workload identity federation instead of secrets
- GitHub Actions with OIDC to Azure; infrastructure as code with Terraform or Bicep
- OpenTelemetry for traces, metrics and logs across .NET services
- Azure AI Search hybrid retrieval (keyword + vector + semantic ranking) for grounded AI
- Structured outputs (JSON schema) for every model response consumed by code
Trial
Promising and production-capable, used where the benefit justifies some risk.
- Microsoft Agent Framework for .NET and Python agent orchestration
- .NET Aspire for local orchestration and service defaults in distributed .NET applications
- Model Context Protocol (MCP) servers to expose enterprise tools to agents, with server-side authorisation
- Azure API Management AI gateway policies for token limits, load balancing and semantic caching
- LangGraph for long-running, stateful agent workflows in Python
- Microsoft Fabric for consolidated analytics and lakehouse workloads
Assess
Worth understanding; we are evaluating where they fit.
- Small language models (such as Phi) for on-device or low-cost, high-volume tasks
- Private inference with vLLM and quantised open-weight models for data-sensitive workloads
- Native AOT for ASP.NET Core APIs where cold-start and memory footprint matter
- Reinforcement and preference fine-tuning for narrow, high-volume tasks
- Automated LLM-as-judge evaluation, calibrated against human review
Hold
Proceed with caution, or avoid for new work.
- Long-lived client secrets and storage account keys in CI/CD pipelines
- The Azure Functions .NET in-process model for new work; use the isolated worker model
- Manually configured cloud infrastructure that is not captured as code
- Autonomous agents with unrestricted write access to production systems
- Fine-tuning before prompting and retrieval have been properly evaluated
- Choosing the largest available model by default without measuring quality, latency and cost
Contact: contact@nexaarislabs.com