Thinking out loud.
Engineering notes, platform updates, and thinking out loud.
Pressure as a platform primitive: one ResourceBroker for LLMs and HTTP
Rate-limiting LLMs and rate-limiting webhooks look like different problems. They are the same problem with different exhaustion strategies. One broker, two methods, one config plane.
Read post →Public template, private infra: where deployment lives for an OSS product
An OSS product has to be both deployable by anyone and operable by the vendor. Those are two artifacts with different properties. One repo either leaks secrets or makes the reference unusable. The fix is two.
Read post →Affected is a platform mode, not a build optimization
Affected detection is usually framed as 'skip rebuilding what didn't change'. Its real value is letting each layer of the stack decide whether to do work at all. Those decisions compound.
Read post →Building a workflow engine that treats humans and agents the same
Most workflow engines bolt human steps on as a special case. We built one where approvals and agent decisions are the same type — same execution model, same state machine.
Read post →Contract-based adapters: swap Kafka for NATS without touching business logic
12 core contracts, strict layer rules, and a dependency graph that physically prevents plugins from importing adapters. How we made infrastructure genuinely swappable.
Read post →Duck typing for plugins: why we killed the plugin interface
No base class. No implements. No runtime registration call. If a package has a manifest with schema: kb.plugin/3, it's a plugin. Here's why that works better than an interface.
Read post →