LAYA AI vs Jev: what changes when you run it locally
A documentation and source review of LAYA next to Jev: deployment choice, benchmark caveats, the router's default checkpoint selection and what a real evaluation would need.
Read articleHi, I’m Sebastian Bimbi.
I’m a senior developer writing about AI, software and the decisions behind what we build.
Featured article
Latest writing
View all writingThe coding agent is becoming a runtime
How GitHub and JetBrains describe the shift from line-by-line suggestions to an agent that owns a scoped task, with permissions, evidence and a stopping point.
MCP's stateless turn
The Model Context Protocol's stateless core removes session affinity from the transport layer. What that changes for scaling, tracing and existing integrations.
The sandbox is part of the product
What makes a coding agent safe enough to run on a real repository: sandbox boundaries, approval policy, network controls and telemetry as separate, layered controls.
Long-horizon coding changes benchmarking
Why a single prompt-latency number can't describe an agentic coding workload, and what MLCommons' agentic-inference work measures instead: trajectories, not calls.
What is Jev? TypeSafe’s decision model, explained
A developer’s guide to Jev, TypeSafe’s model for typed decisions, and the difference between a valid answer and a correct one.
Jev support triage with Choice, Score and Noul
Design a support workflow with Choice, Score and Noul. Inspect an offline-tested example that keeps every suggestion under human review.
Evaluating Jev before production
A practical plan to evaluate Jev: meaningful test cases, confidence, error costs and the work that remains with people.
A few things on my mind
How much of a feature’s behaviour should live in typed code, and how much in a model’s judgement? That question sits behind most of what I’m writing about Jev right now, and I don’t expect to answer it once and be done.
Shorter notes are next: the archive starts empty and fills in with whatever I’m working through between articles. See Notes
Your application context
ChoiceScoreNoul
Your code decides what happens