prisml / demo

Predictions that
ship with your code.

A real ONNX model running in-process. Adjust the inputs, see the latency, then break the schema.

artifact userChurn.onnx
algorithm gbm
task binary_classification
compiled 3/10/2026

Live inference

Adjust inputs. Get a real prediction.

Each submit hits /api/demo/predict, which calls session.predict() synchronously in the Node.js route. The latency shown is actual wall-clock ONNX inference time.

Feature inputs

Churn risk model

34
0d90d
192
$0$1,000
7
015

Prediction output

Will churn

1 — positive
Raw output
1
Inference latency
20 ms
Schema hash
a3f8c1b2
Artifact
userChurn.onnx

Latency comparison

Why in-process matters

PrisML in-process 20 ms
HTTP microservice ~120 ms
External inference API ~400 ms
Python subprocess ~800 ms

Scale: 0 – 800 ms (Python subprocess = 100%). PrisML runs ONNX Runtime in-process — no serialisation, no network hop.

Schema drift guard

The model knows what schema it was trained on.

Every artifact carries a SHA-256 hash of the Prisma schema at compile time. If the schema changes and you try to initialise predictions, PredictionSession throws a typed SchemaDriftError before a single inference runs. No silent incorrectness.

$ ls -lh .prisml/

-rw-r--r--52KuserChurn.onnx

-rw-r--r-- 4KuserChurn.metadata.json

# both files are bound to schema hash a3f8c1b2

Simulate schema change

Toggle a schema migration

This artifact was compiled against schema hash a3f8c1b2. Simulate adding a field to the User model and watch the session refuse to proceed.

Session response

Prediction runs normally

OK 
// Schema hash matches — inference proceeds normally.

const result = await session.predict(churnModel, user);
// result.prediction → "0"
// result.timestamp  → "2026-03-08T12:00:00.000Z"

The session holds a3f8c1b2 — the hash of the schema at compile time. When calling session.load(), PrisML re-hashes the current schema file and compares. A match means the feature resolvers are still valid against the live data shape.

What the error looks like

SchemaDriftError in practice

// After adding a new field to schema.prisma:
// model User {
//   ...
//   subscriptionTier String   // ← new field
// }

// session.load() re-hashes the current schema file before loading.
await session.load(churnModel); // schema now hashes to 7d2e9f1a

// ✖ Throws:
// SchemaDriftError: Model 'userChurn' was trained on schema
// hash a3f8c1b2 but the current schema hash is 7d2e9f1a.
// Run `prisml train` to rebuild the artifact.

Integration

Three files. Zero infrastructure.

Define a model in TypeScript, compile once at build time, predict in-process at runtime. The same feature resolvers run during training and inference — the behaviour cannot drift. 

import { defineModel } from '@vncsleal/prisml';

export const churnModel = defineModel<User>({
  name: 'userChurn',
  modelName: 'User',
  output: {
    field: 'willChurn',
    taskType: 'binary_classification',
    resolver: (u) => u.willChurn,
  },
  features: {
    daysSinceActive: (u) =>
      Math.floor((Date.now() - u.lastActiveAt.getTime()) / 86_400_000),
    monthlySpend: (u) => u.monthlySpend,
    supportTickets: (u) => u.supportTickets,
  },
  algorithm: { name: 'gbm' },
  qualityGates: [
    { metric: 'f1', threshold: 0.80, comparison: 'gte' },
    { metric: 'accuracy', threshold: 0.85, comparison: 'gte' },
  ],
});
$ prisml train

◉ Loading Prisma schema...
✔ Schema loaded (hash: a3f8c1b2)
◉ Loading model definitions...
✔ Loaded 1 model definition(s)
◉ Validating models...
✔ Models validated
◉ Extracting training data via Prisma...
◉ Training userChurn (gbm)...
✔ Training complete
◉ Writing artifacts...
✔ Artifacts written to ./.prisml

[OK] Training complete

Artifacts:
  userChurn.metadata.json
  userChurn.onnx
import { PredictionSession } from '@vncsleal/prisml';
import { churnModel } from './models/churn';

const session = new PredictionSession();
await session.load(churnModel); // resolves .prisml/ and prisma/schema.prisma

const result = await session.predict(churnModel, user);

console.log(result.prediction);  // "1" or "0"
console.log(result.timestamp);   // "2026-03-08T12:00:00.000Z"
// Inference ran in ~0.4ms — in the same V8 process.
$ npm install @vncsleal/prisml
View source on GitHub