A reason to build.
Infrastructure knowledge was scattered across notes and services. An assistant needed enough context to answer useful questions without receiving unrestricted control over production.
How the pieces fit.
- 01
Connected the assistant to a shared knowledge base through MCP, so architectural decisions and operational notes travel with the conversation.
- 02
Added a constrained observation layer for service and host inspection. Read-only command rules and secret-aware file access define the boundary.
- 03
Connected voice messages to a private, CPU-based Whisper transcription service, removing the dependency on cloud transcription credits.
- 04
Allowed the assistant to write plans and discoveries back into the knowledge base. Changes to hosts still happen through a separate execution path.
What came out of it.
Deployed an operational gateway that can retrieve knowledge, observe systems, and preserve context without being allowed to mutate hosts. Service health and live inference were verified during deployment.
Documented deployment, health checks, adapter verification, and live inference checks. Private source; this case study describes the design without exposing operational details.
What I’d carry forward.
An agent’s useful scope and its execution authority are separate design decisions. A working health endpoint also isn’t enough: the transcription adapter and actual inference path need their own checks.
Working on something with similar edges?
Let’s compare notes ↗