A small learned decision runtime, built in Rust. Roughly 100K parameters, with the execution path treated as a first-class engineering problem.
How much decision-making can you fit into a few microseconds?
The interesting problem
A model is only one part of a decision system. Getting inputs ready, executing the model and returning a usable result all sit inside the same latency budget. At this scale, the machinery around inference deserves as much attention as the inference itself.
A deliberately small runtime
MICROJEV pairs a roughly 100K parameter model with a Rust runtime. Its defining constraints are a small model and an allocation-free hot path. The point is to make a learned decision fit into a system that cannot afford a long pause.
Reading the result
The reported p50 is 7.4 microseconds, with zero hot-path allocations. Those numbers describe a specific measurement, not a universal guarantee. A useful comparison also needs the workload, hardware and benchmark methodology; a median alone says nothing about the slowest requests.
Reported project figures. Hardware, workload and full benchmark methodology are not published here.