Protocol-first AI Runtime

Safe model training, inference, and self-retraining

A serverless-first AI runtime with explicit contracts, snapshot-before-training enforcement, and complete lineage tracing for reproducible, auditable ML pipelines.

Get Started View on GitHub

100%

Reproducible

Zero

Mutable Training Data

Full

Lineage Tracing

Datasource Types

Core Guarantees

Built from the ground up for safe, reproducible, and auditable ML operations

Protocol-First Contracts

Explicit versioned envelopes like pulse.source.sql.extract.v1 ensure clear, typed communication across all operations.

Snapshot-Before-Training

Training never runs on mutable live data. Every job uses a versioned dataset snapshot for complete reproducibility.

Full Lineage Tracing

Track every prediction back to its source data, model version, and training configuration with complete audit trails.

Self-Retraining

Automatic retraining decisions driven by drift detection, feedback signals, data freshness, and scheduled policies.

Read-Only Datasources

Encrypted connection configs with allowlisted SQL objects. No arbitrary queries—only predefined, safe extractions.

Serverless-First

Externalized metadata and object storage design. Scale to zero when idle, burst when needed.

Modular Architecture

Clean separation of concerns with explicit boundaries. Each package has a single responsibility and communicates through well-defined interfaces.

api/

HTTP API exposing named protocol operations

worker/

Background orchestration for recovery and retraining

API Layer

POST /mutations/:name

POST /queries/:name

Runtime

Extract

Train

Evaluate

Promote

Predict

Storage

Metadata Store

Object Store

Protocol Examples

Every operation follows explicit, versioned contracts with idempotency keys and full audit metadata

json

{
  "protocol": "pulse",
  "kind": "mutation",
  "name": "pulse.datasource.register.v1",
  "version": "v1",
  "messageId": "msg_123",
  "timestamp": "2026-03-26T00:00:00.000Z",
  "idempotencyKey": "datasource-tenant-a-orders",
  "tenantId": "tenant_a",
  "payload": {
    "name": "Orders Warehouse",
    "type": "postgres",
    "connectionConfig": {
      "dsn": "postgres://readonly@warehouse/orders"
    },
    "allowedSchemas": ["public"],
    "allowedViews": ["ai_orders_v1"]
  }
}

Extract Snapshot

json

{
  "protocol": "pulse",
  "kind": "mutation",
  "name": "pulse.source.sql.extract.v1",
  "version": "v1",
  "messageId": "msg_124",
  "timestamp": "2026-03-26T00:00:01.000Z",
  "tenantId": "tenant_a",
  "payload": {
    "dataSourceId": "ds_...",
    "queryDefinition": {
      "kind": "view",
      "identifier": "ai_orders_v1",
      "cursorField": "updated_at"
    },
    "extractionMode": "full_snapshot",
    "chunkSize": 500
  }
}

Trace Lineage

json

{
  "protocol": "pulse",
  "kind": "query",
  "name": "pulse.lineage.trace.v1",
  "version": "v1",
  "messageId": "msg_125",
  "timestamp": "2026-03-26T00:00:02.000Z",
  "tenantId": "tenant_a",
  "payload": {
    "entityId": "pred_...",
    "entityType": "prediction"
  }
}

Pluggable Datasources

Connect to your existing data infrastructure with secure, read-only connectors that enforce allowlists and encryption by default.

PostgreSQL

Full support with schema inspection and incremental extraction

MySQL

Read-only connections with allowlisted views and tables

SQL Server

Enterprise database support with cursor-based pagination

Object Storage

S3-compatible storage for snapshots and model artifacts

HTTP

REST API sources (coming soon)

Connector Interface

Every connector implements a standard interface for consistent behavior across all datasource types.

inspectSchema()

validateConnection()

extractSnapshot()

streamData()

estimateSize()

Ready to build safe ML pipelines?

Get started with Pulse in minutes. Clone the repository, run the API, and start building reproducible, auditable AI systems.

Get Started Read the Docs

Quick Start

# Clone and install
git clone https://github.com/DiogoAngelim/pulse
cd pulse
npm install

# Start the API and worker
npm run start:api
npm run start:worker