Laya on Railway
Open-source decision engine API with calibrated probabilities from one forward pass, on Railway.
Laya answers typed questions about a piece of text, such as which option fits, where it sits on a scale or yes or no, and returns calibrated probabilities from a single forward pass instead of generated text. It runs on CPU and covers English and more than 100 other languages.
It ships laya-serve, an HTTP server that speaks the request and answer format of TypeSafe's Jev decision API, so code written against that API can point its base URL at a Laya deployment. Typical uses are routing tickets and emails, policy or relevance checks around an LLM call, and rating urgency or sentiment.
Laya's code and weights are Apache-2.0.
What Laya can do
- Typed decisions: option choice, scale rating and yes or no, with calibrated probabilities
- Endpoints /v1/systemone, /v1/systemone/batch (up to 64 texts) and /v1/route
- English (ModernBERT-large) and multilingual (mmBERT-base) checkpoints
- CPU-only inference
- Compatible with the Jev /v1/systemone request format
- API key required
What people use it for
- Triage tickets, emails and chats into queues, with low-confidence ones handed to a person
- Policy, safety or relevance checks around an LLM call
- Rating urgency, sentiment or lead quality
- A self-hosted endpoint for Jev API clients
Videos
Running Laya on Railway
The Railway template builds Laya 0.4.28 with Python packages pinned by SHA-256 and CPU-only PyTorch, and runs laya-serve as one public service with a /health check. It refuses to start without a generated LAYA_API_KEY. Checkpoints are downloaded from Hugging Face on first boot, pinned to a reviewed revision, verified by SHA-256 and kept on a volume at /data. Memory drives cost: about 5 GB of RAM was measured on Railway with both checkpoints loaded.
Laya templates on Railway
Tracklayer templates are listed first. Other Laya templates follow, ordered by total deployments on Railway. How this works.
Deployment numbers as of 2026-11-20.
Questions about Laya
How do I call it?
Send requests to your domain with the generated LAYA_API_KEY from the service variables. The endpoints are /v1/systemone, /v1/systemone/batch and /v1/route, using the Jev decision API format.
How much memory does it need?
About 5 GB of RAM was measured on Railway with both default checkpoints loaded, so the Trial plan's 1 GB is too small. A multilingual-only setup is documented at roughly 2 GB.
Is the model kept across redeploys?
Yes. Checkpoints are cached on a volume at /data, about 1.5 GB for the defaults, and downloaded from Hugging Face on first boot.
Is it open to anyone who finds the URL?
No. Upstream's server is open by default, but the template refuses to start without an API key.
Is Laya the same as Jev?
No. Laya is a separate open-source project that speaks Jev's request format so Jev clients can use it.
Similar apps
Tracklayer maintains the templates marked as such; Laya is a trademark of its owners and this site is not affiliated with them.