Deploying models

What ships inside a package

Every dataset the AI Factory publishes contains a PROVENANCE.json alongside its artifacts, so a file that has been copied onto a workstation can still answer where it came from:

• the listing slug, title, licence and marketplace URL it was published under
• a pseudonymous creator id — stable across that creator’s packages, not reversible to an account, and never accompanied by an email or name
• every source clip it was built from, with the SHA-256 of the bytes actually shipped, plus a SHA-256 per generated artifact
• the screening verdict, which flag categories fired, and which categories of personal data were stripped from the prompt before drafting — the categories, never the values
• an explicit list of what the record does not contain, so its absence reads as a decision rather than an oversight

The same record is what a signed .deb will embed under /usr/share/doc when Jetson packaging ships. The manifest is format-agnostic on purpose.

Deployment surfaces

• Hosted — managed endpoint with autoscaling and a signed invoke URL
• Jetson — a signed .deb package with systemd unit and Triton runtime
• ROS 2 — drop-in launch file that wires inputs and outputs to topics
• Self-hosted container — the nodedata/runtime image runs anywhere

Hosted inference

Create a deployment from any model you have licensed. Cold start is typically under three seconds for models under 2 GB.

deployment = node.deployments.create(
    model="acme/grasp-policy-v2",
    revision="1.4.0",
    accelerator="nvidia-l4",        # nvidia-l4 | nvidia-a100 | nvidia-h100 | cpu
    min_replicas=1,
    max_replicas=8,
    scale_to_zero=True,
)

print(deployment.endpoint)
# → https://infer.nodedata.dev/d/dpl_01HZ.../invoke

Invoke

curl https://infer.nodedata.dev/d/dpl_01HZ.../invoke \
  -H "Authorization: Bearer $NODE_DATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "inputs": { "rgb": "<base64>", "depth": "<base64>" }
  }'

Jetson and edge packages

Node Factory generates a Jetson-ready package containing a quantized engine plan, a systemd unit, and the inference server. Packages are signed and verified at install time.

# Generate a Jetson-ready package for an asset
node-data export \
  --model acme/grasp-policy-v2 \
  --revision 1.4.0 \
  --target jetson-orin \
  --precision fp16 \
  --out ./pkg

# On the device
sudo dpkg -i ./pkg/nodedata-grasp-policy_1.4.0_arm64.deb
systemctl enable --now nodedata-grasp-policy

ROS 2

The nodedata_runtime ROS package exposes any model as a node with configurable input and output topics. It supports both rclpy and rclcpp.

# launch file
ros2 launch nodedata_runtime model.launch.py \
  asset:=acme/grasp-policy-v2 \
  revision:=1.4.0 \
  topic_in:=/camera/rgb \
  topic_out:=/policy/action

Self-hosted container

FROM nodedata/runtime:1
ENV NODE_DATA_API_KEY=<provisioned-at-deploy>
RUN nodedata pull acme/grasp-policy-v2 --revision 1.4.0
CMD ["nodedata", "serve", "--port", "8080"]

Precision and quantization

Logs and metrics

• Structured logs streamed via GET /deployments/{id}/logs
• p50 / p95 / p99 latency and throughput on the deployment dashboard
• OpenTelemetry exporter for shipping traces to your own backend