From trained blob to deployed model in minutes

dstk is a data science toolkit for model registries, deployment, and monitoring. Infrastructure as Code for your ML models, in a way that mostly doesn't suck.

Get started Read the docs

$pipinstallgit+https://github.com/dstk-labs/pydstk.git

python3 deploy.py

from pydstk import ModelRegistry, Deployer

# Register your model
registry = ModelRegistry()
model = registry.register(
    name="fraud-detector",
    artifact="./model.pkl",
    version="1.2.0"
)

# Deploy with one call
Deployer().deploy(model)

✓ Registered fraud-detector v1.2.0
✓ Deployed to https://models.acme.io/fraud-detector
✓ Inference API live (200 OK)

Plays nice with

scikit-learn
PyTorch
TensorFlow
XGBoost
HuggingFace
ONNX

How it works

Register. Deploy. Monitor.

Three concepts, minimal configuration, and you pick the interface: Python, CLI, or Web UI.

Step 1: Register

Immutable model registry

Point dstk at your trained artifact. It gets versioned and stored immutably. No more "which pickle file was that?"

model = registry.register(
  name="churn-predictor",
  artifact="./model.pkl"
)

Step 2: Deploy

Painless, hardware agnostic

dstk wraps your model in a container with a standard inference API. No Dockerfiles, no Kubernetes YAML, no tears.

Deployer().deploy(model)

# That's literally it

Step 3: Monitor

Full audit trail, zero config

Every deployment is tracked. See what's running, when it shipped, which version, and roll back if something goes wrong.

# From Python
registry.deployments("churn-predictor")

[
  { version: "1.2.0", status: "live" },
  { version: "1.1.0", status: "retired" }
]

# Or from the CLI
$ pydstk deployments list churn-predictor

VERSION  STATUS   DEPLOYED
1.2.0    live     2 hours ago
1.1.0    retired  3 days ago
1.0.0    retired  2 weeks ago

Comparisons

Why not just use MLflow?

Existing tools are either too bloated, too proprietary, or too painful. Here's how dstk stacks up.

Capability	dstk	MLflow	SageMaker	Azure ML
Open source	Yes	Yes	No	No
Immutable model registry	Yes	Partial	Yes	Yes
One command deploy	Yes	No	Partial	Partial
Hardware agnostic	Yes	Yes	No	No
Minimal configuration	Yes	No	No	No
Standard inference API	Yes	Partial	Yes	Yes
No vendor lock in	Yes	Yes	No	No
Doesn't suck	Yes	Partial	Partial	Partial