Notes

What we learned on a specific project, written up so it’s useful on yours.

A pattern that held, a failure worth naming, a decision and why it went the way it did.

  1. 01generative AI & model serving16 Jul 2026Region-pinning Claude on Vertex AI without fighting your model registryClaude-on-Vertex serves from only a few regions; rarely the one hosting your agent. Pin conditionally, and return a model object so the registry's regex resolution gets out of the way.
  2. 02agentic AI & platform engineering09 Jul 2026A paved road for LLM agents: golden paths, resilient tools, boring deploysTreat agent delivery like any other production system; a scaffold-to-deploy golden path, tool connections that degrade instead of crashing the turn, and secrets tied to per-agent identity.
  3. 03networking & security23 Jun 2026Fail closed on database TLS: the driver default is cleartextMost drivers default to no encryption and managed databases rarely force it, so a missing connection-string parameter connects in the clear with no error; refuse to boot without verified TLS, and prove it by reading the negotiated cipher.
  4. 04configuration & platform data20 May 2026Honest per-environment scoring: when a glob eats your sandboxA `**/production/**` glob that quietly swallows a sandbox nested under production, and why you should compute the criteria that don't vary once instead of faking N independent audits.
  5. 05containers, clustering & networking04 Nov 2025Ephemeral jobs as transport: probing from inside a cluster without a resident agentWhen a reachability test only means something from the real source network, ship the probe as a throwaway Kubernetes Job and read the answer out of the pod's logs.
  6. 06data & observability20 Oct 2025A window is not a ledger: metrics that survive a rolling-window feedWhy "total raised" is uncomputable from a feed that only ever shows you the last 31 days, and what to report instead.