layajev

Run a Jev-compatible decision API on your own machine. layajev serves a verified Laya model through a small HTTP API: send a state and questions, and receive typed answers. Inference stays local and in-process—no hosted inference service or Python server is involved.

layajev is a compatibility subset, not the Jev model or a promise of identical decisions. See the supported request forms and errors.

• Easy to call: GET /v1/models and POST /v1/systemone return structured answers for supported choice, score, and noul questions.
• Local by design: the server runs inference in its own process; serving does not contact a model host.
• Explicit model lifecycle: fetch and conversion are separate, operator-invoked steps, and serving opens a verified bundle.

Start with npx

On a supported linux/amd64 host, check the packaged native runtime, then serve an existing verified bundle:

npx -y @metalagman/layajev@latest doctor
npx -y @metalagman/layajev@latest serve --bundle /absolute/path/to/verified-bundle

The API listens on 127.0.0.1:8080 by default. The npm package includes the native runtime, not model weights. doctor checks the runtime but does not prepare a model. If you do not have a bundle yet, follow the from-zero model setup; it fetches the pinned official source, verifies it, and converts it explicitly.

Ask the running API a question:

curl -sS http://127.0.0.1:8080/v1/models

curl -sS http://127.0.0.1:8080/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"model":"laya-multilingual","state":"I was charged twice.","questions":{"billing":{"type":"noul","instructions":"Is this about billing?"}}}'

Use the model name returned by /v1/models if it differs from laya-multilingual. Choice, score, and noul questions are supported within the documented API subset.

Run it your way

• npx: the quickest way to run doctor, fetch, or serve on linux/amd64. CLI and model setup.
• Docker: build a production image that contains its own verified model bundle; no model mount is needed at runtime. Container and Compose guide.
• Go source: build or modify the adapter in this repository. Contributing guide.

serve never downloads or converts a model. The server has no built-in TLS or authentication: keep the default loopback listener, or place it behind an authenticating reverse proxy before allowing remote traffic. Only the pinned linux/amd64 native candidate is qualified; other platforms are not advertised as supported.

layajev is MIT-licensed. Model and SDK licenses are separate and must be reviewed before redistribution. The reusable inference library lives in laya-go.