PrisML

Compiler-first machine learning for TypeScript + Prisma.

PrisML is for developers who want a narrow, local, reviewable ML workflow:

• define models in TypeScript
• train them at build time
• commit immutable ONNX + metadata artifacts
• run predictions in process with schema checks

It is not a hosted ML platform, an online learning system, or a runtime experimentation layer.

Who It Is For

PrisML is aimed at TypeScript and Prisma teams that want:

• deterministic builds
• schema-aware model artifacts
• local inference without adding a separate prediction service
• an ML workflow that can live inside the same repo as application code

Current Promise

The current package provides:

• defineModel() for typed model definitions
• prisml train for build-time training
• prisml check for schema-only validation
• PredictionSession for runtime loading and inference
• immutable model.onnx + model.metadata.json artifacts

Quick Start

Install the package:

npm install @vncsleal/prisml

Define a model in prisml.config.ts:

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

export const userChurnModel = defineModel<User>({
  name: 'userChurn',
  modelName: 'User',
  output: {
    field: 'churned',
    taskType: 'binary_classification',
    resolver: (user) => user.churned,
  },
  features: {
    loginCount: (user) => user.loginCount,
    plan: (user) => user.plan,
    daysSinceSignup: (user) =>
      Math.floor((Date.now() - user.createdAt.getTime()) / 86400000),
  },
  algorithm: {
    name: 'forest',
    version: '1.0.0',
  },
});

Train artifacts:

npx prisml train --config ./prisml.config.ts --schema ./prisma/schema.prisma

Run predictions:

import { PredictionSession } from '@vncsleal/prisml';
import { userChurnModel } from './prisml.config';

const session = new PredictionSession();
await session.load(userChurnModel);

const result = await session.predict(userChurnModel, user);
console.log(result.prediction);

Docs

• ROADMAP.md - direction, milestones, OSS path, monetization path
• MANIFEST.md - project stance and values
• docs/ARCHITECTURE.md - technical boundaries and invariants
• docs/GUIDE.md - usage guide
• docs/API.md - API reference
• docs/INDEX.md - documentation map

Development

pnpm install
pnpm --dir packages/prisml test
pnpm --dir packages/prisml build

License

MIT