It took me 18 hours to add Jev

It took me 18 months to build The Pipeline Framework.

18 days to turn it into an AI operating system.

18 hours to add an entirely new kind of AI.

TypeSafe released Jev only days ago: a System One model that doesn’t generate language. It makes fast, bounded, probabilistic decisions.

TPF had been built around functional pipelines, "Functional Core, Imperative Shell", reactiveness, scalability, then generative LLMs, RAG, tools, agent loops, MCP, human interaction, durable effects...

So how much architecture would a fundamentally different model require? Turns out: almost none. So Jev became another Query provider.

invoice-judgment-model:
  provider: decision.query.jev
  version: 1
  config:
    model: typesafe/jev-1.13

I added a provider-neutral decision protocol:

DecisionRequest → DecisionResult

with bounded Choice, Noul and Score questions, then plugged Jev underneath it.

A real Java application was using it the same day — despite Jev not yet having an official Java SDK.

And I think that is the milestone I’m happiest about: not that TPF supports the latest AI model, but that it didn’t particularly care that it was new.

An LLM is a Query. RAG retrieval is a Query. A database lookup is, you guessed it, also a Query. So now a System One probabilistic decision is a Query, as it should.

Different computational models. Different latency profiles. Different semantics at the provider level. Same application architecture.

The first 18 months were spent finding the abstractions.

The next 18 days exploited them.

The next 18 hours barely disturbed them.

That’s the compounding return of solid architecture.