Skip to content

Research spikes

Every research spike behind the architecture decisions, in order. Each grounds its findings in a first-party Microsoft Learn source or a named vendor's own documentation, and each traces forward to one or more ADRs.

SpikeTopic
SPIKE-01MAI-Image-2.5 image model
SPIKE-02MAI-Voice-2 text-to-speech model
SPIKE-03Tenant and subscription readiness
SPIKE-04Identity, secrets, security, and responsible AI
SPIKE-05Cost model and governance
SPIKE-06Publish-pipeline integration
SPIKE-07Speech models beyond MAI and native Azure (word-sync and lip-sync)
SPIKE-08Foundry Local (on-device inferencing)
SPIKE-09AI Foundry model workloads on Azure Local (Arc-connected, cluster-scale)
SPIKE-10Latest available in-tenant GPT model
SPIKE-11A newer Grok than grok-4-1-fast-reasoning
SPIKE-12Broader image, animation, and video generation alternatives
SPIKE-13Tenant-wide voice/TTS survey
SPIKE-14Tenant-wide model and region survey
SPIKE-15Niche and emerging reviewer models for code and document review
SPIKE-16Photorealistic virtual-trainer avatar
SPIKE-17Governing agent MCP tools with an APIM AI gateway
SPIKE-18Foundry Local on Windows Server, and whether the Arc run-command can install it (track 2)
SPIKE-19Where the ARM and Kubernetes seam sits for Foundry Local on Azure Local (track 3)
SPIKE-20Cost-first Azure Monitor, Foundry, Grafana, and hybrid observability
SPIKE-21Solution observability for Azure AI Foundry
SPIKE-22The Foundry Local model catalog, and whether tracks 2 and 3 draw from one catalog or two
SPIKE-23Foundry Local install artifacts, and the real mechanics of Azure Arc run command
SPIKE-25Hardware sizing and capacity planning for the two local Foundry tracks
SPIKE-26The cost model for the two local tracks, and where a spend cap can actually be enforced
SPIKE-27What Azure can and cannot see for the local deployment tracks
SPIKE-28Networking, ingress, TLS, and storage for Foundry Local on Azure Local
SPIKE-29Lifecycle, upgrade, and drift for the two Foundry Local tracks
SPIKE-31Cross-track capability and feature parity across the three deployment targets
SPIKE-32Every model against every region, and what actually differs between them
SPIKE-33Which client tools can point at this endpoint, and what each one gets wrong
SPIKE-34Orchestration options, and which of them can be pointed at this account