Publishing models as APIs

Lecture 15

Dr. Benjamin Soltoff

Cornell University
INFO 4940/5940 - Fall 2025

October 21, 2025

Announcements

Updates to Posit Workbench
Project 01 EDA due Thursday

Learning objectives

Create bundled model objects that can be saved to disk
Implement versioning for model objects
Review application programming interfaces (APIs)
Generate a REST API for a model using Vetiver

Application exercise

`ae-14`

Instructions

Go to the course GitHub org and find your ae-14 (repo name will be suffixed with your GitHub name).
Clone the repo in Positron, run renv::restore() to install the required packages, open the Quarto document in the repo, and follow along and complete the exercises.
Render, commit, and push your edits by the AE deadline – end of the day

MLOps

MLOps with {vetiver}

Vetiver, the oil of tranquility, is used as a stabilizing ingredient in perfumery to preserve more volatile fragrances.

If you develop a model…

you can operationalize that model!

If you develop a model…

you likely should be the one to operationalize that model!

Tompkins County housing data

Home sale prices for Tompkins County, NY between 2022-23
Can certain measurements be used to predict the sale price?
Data collected from Redfin

Tompkins County housing data

R

housing <- read_csv("data/tompkins-home-sales.csv")

Python

import pandas as pd
housing = pd.read_csv('data/tompkins-home-sales.csv')

N = 1,225
A numeric outcome, price
Other variables to use for prediction:
- beds, baths, area, and year_built are numeric predictors
- town and municipality could be nominal predictors
- sold_date could be a date predictor

Home prices in Tompkins County, NY

sold_date	price	beds	baths	area	lot_size	year_built	hoa_month	town	municipality	long	lat
2022-08-16	335500	3	2.0	1957	4.50000000	1880	NA	Ulysses	Unincorporated	-76.67680	42.53255
2022-11-14	331500	3	2.0	1416	0.58999082	1930	NA	Lansing	Unincorporated	-76.50347	42.53340
2022-03-31	302385	3	1.5	1476	0.20000000	1900	NA	Ithaca	Ithaca city	-76.50439	42.44250
2022-09-28	285000	3	2.0	1728	0.46999541	2002	NA	Dryden	Dryden village	-76.29495	42.48415
2022-07-22	350000	4	1.0	1698	0.12396694	1925	NA	Ithaca	Ithaca city	-76.50146	42.43264
2023-11-28	225000	2	1.5	1047	0.08000459	1939	NA	Ithaca	Ithaca city	-76.50576	42.43373
2023-09-13	285000	3	2.0	2311	1.26999541	1965	NA	Caroline	Unincorporated	-76.33375	42.39048
2023-06-23	145000	2	2.0	1215	0.03999082	1990	NA	Danby	Unincorporated	-76.49228	42.38340
2023-11-27	90900	5	3.0	2238	0.38000459	1880	NA	Groton	Groton village	-76.36311	42.58533
2022-11-09	467500	6	4.0	2304	0.13000459	2017	NA	Ithaca	Ithaca city	-76.50205	42.43136

Time for building a model!

Spend your data budget

R

library(tidymodels)
set.seed(123)

housing_split <- housing |>
  mutate(price = log10(price)) |>
  initial_split(prop = 0.8)

housing_train <- training(housing_split)
housing_test <- testing(housing_split)

Python

from sklearn import model_selection
import numpy as np
np.random.seed(123)
X, y = housing[["beds", "baths", "area", "year_built"]], np.log10(housing["price"])
X_train, X_test, y_train, y_test = model_selection.train_test_split(
    X, y,
    test_size = 0.2
)

Fit a linear regression model 🚀

Or your model of choice!

R

housing_fit <- workflow(
  price ~ beds + baths + area + year_built,
  linear_reg()
) |>
  fit(data = housing_train)

Python

from sklearn import linear_model
housing_fit = linear_model.LinearRegression().fit(X_train, y_train)

📝 Fit a model

Instructions

Split your data in training and testing.

Fit a model to your training data.

03:00

Create a deployable bundle

Deploy preprocessors and models together

Create a deployable model object

R

library(vetiver)
v <- vetiver_model(housing_fit, "tompkins-housing")
v


── tompkins-housing ─ <bundled_workflow> model for deployment 
A lm regression modeling workflow using 4 features

Python

from vetiver import VetiverModel
v = VetiverModel(housing_fit, "tompkins-housing", prototype_data = X_train)
v.description

'A scikit-learn LinearRegression model'

Deploy preprocessors and models together

What is wrong with this?

📝 Create a vetiver model object

Instructions

Create your vetiver model object.

Check out the default description that is created, and try out using a custom description.

Show your custom description to your neighbor.

05:00

Version your model

pins 📌

The pins package publishes data, models, and other R and Python objects, making it easy to share them across projects and with your colleagues.

You can pin objects to a variety of pin boards, including:

a local folder (like a network drive or even a temporary directory)
Posit Connect
Amazon S3
Azure Storage
Google Cloud

library(pins)

board <- board_temp()
board |> vetiver_pin_write(v)

Creating new version '20251022T155624Z-64a74'
Writing to pin 'tompkins-housing'

Create a Model Card for your published model
• Model Cards provide a framework for transparent, responsible reporting
• Use the vetiver `.Rmd` template as a place to start

from pins import board_temp
from vetiver import vetiver_pin_write

board = board_temp(allow_pickle_read = True)
vetiver_pin_write(board, v)

Model Cards provide a framework for transparent, responsible reporting. 
 Use the vetiver `.qmd` Quarto template as a place to start, 
 with vetiver.model_card()
Writing pin:
Name: 'tompkins-housing'
Version: 20251022T115624Z-1ee78

📝 Pin your model

Instructions

Pin your vetiver model object to a temporary board.

Retrieve the model metadata with pin_meta().

05:00

Version your model

Fit a random forest

R
Python

rf_rec <- recipe(
  price ~ beds + baths + area + year_built + town,
  data = housing_train
) |>
  step_impute_mean(all_numeric_predictors()) |>
  step_impute_mode(all_nominal_predictors())

housing_fit <- workflow() |>
  add_recipe(rf_rec) |>
  add_model(rand_forest(trees = 200, mode = "regression")) |>
  fit(data = housing_train)

from sklearn.ensemble import RandomForestRegressor
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline

# Define feature columns
numeric_features = ["beds", "baths", "area", "year_built"]
categorical_features = ["town"]

# Create preprocessing steps
numeric_transformer = SimpleImputer(strategy="mean")
categorical_transformer = Pipeline(steps=[
    ("imputer", SimpleImputer(strategy="most_frequent")),
    ("onehot", OneHotEncoder(handle_unknown="ignore"))
])

# Combine preprocessors
preprocessor = ColumnTransformer(
    transformers=[
        ("num", numeric_transformer, numeric_features),
        ("cat", categorical_transformer, categorical_features)
    ]
)

# Create pipeline with preprocessor and model
housing_fit = Pipeline(steps=[
    ("preprocessor", preprocessor),
    ("regressor", RandomForestRegressor(n_estimators=200, random_state=123))
])

# Prepare training data with all features
X_train_full = housing.loc[X_train.index, numeric_features + categorical_features]
housing_fit.fit(X_train_full, y_train)

Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('num', SimpleImputer(),
                                                  ['beds', 'baths', 'area',
                                                   'year_built']),
                                                 ('cat',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(strategy='most_frequent')),
                                                                  ('onehot',
                                                                   OneHotEncoder(handle_unknown='ignore'))]),
                                                  ['town'])])),
                ('regressor',
                 RandomForestRegressor(n_estimators=200, random_state=123))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Pipeline

?Documentation for PipelineiFitted

Parameters

	steps	[('preprocessor', ...), ('regressor', ...)]
	transform_input	None
	memory	None
	verbose	False

preprocessor: ColumnTransformer

?Documentation for preprocessor: ColumnTransformer

Parameters

	transformers	[('num', ...), ('cat', ...)]
	remainder	'drop'
	sparse_threshold	0.3
	n_jobs	None
	transformer_weights	None
	verbose	False
	verbose_feature_names_out	True
	force_int_remainder_cols	'deprecated'

num

['beds', 'baths', 'area', 'year_built']

SimpleImputer

?Documentation for SimpleImputer

Parameters

	missing_values	nan
	strategy	'mean'
	fill_value	None
	copy	True
	add_indicator	False
	keep_empty_features	False

cat

['town']

SimpleImputer

?Documentation for SimpleImputer

Parameters

	missing_values	nan
	strategy	'most_frequent'
	fill_value	None
	copy	True
	add_indicator	False
	keep_empty_features	False

OneHotEncoder

?Documentation for OneHotEncoder

Parameters

	categories	'auto'
	drop	None
	sparse_output	True
	dtype	<class 'numpy.float64'>
	handle_unknown	'ignore'
	min_frequency	None
	max_categories	None
	feature_name_combiner	'concat'

RandomForestRegressor

?Documentation for RandomForestRegressor

Parameters

	n_estimators	200
	criterion	'squared_error'
	max_depth	None
	min_samples_split	2
	min_samples_leaf	1
	min_weight_fraction_leaf	0.0
	max_features	1.0
	max_leaf_nodes	None
	min_impurity_decrease	0.0
	bootstrap	True
	oob_score	False
	n_jobs	None
	random_state	123
	verbose	0
	warm_start	False
	ccp_alpha	0.0
	max_samples	None
	monotonic_cst	None

Version your model

R

library(pins)
library(vetiver)

board <- board_temp()
v <- vetiver_model(housing_fit, "tompkins-housing", versioned = TRUE)
board |> vetiver_pin_write(v)

Python

from pins import board_temp
from vetiver import vetiver_pin_write

board = board_temp(versioned = True, allow_pickle_read = True)
v = VetiverModel(housing_fit, "tompkins-housing", prototype_data = X_train)
vetiver_pin_write(board, v)

📝 Pin a new version

Instructions

Create a new vetiver model object using your linear regression model that is explicitly versioned = TRUE and pin to your temporary board.

Then train a random forest model and create a new vetiver model object that is also versioned = TRUE with the same name.

Write this new version of your model to the same pin, and see what versions you have with pin_versions().

05:00

Make it easy to do the right thing

Robust and human-friendly checking of new data
Track and document software dependencies of models
Model cards for transparent, responsible reporting

You can deploy your model as a…

REST API

Application programming interface (API)

An interface that can connect applications in a standard way

Representational State Transfer (REST)
Uniform Resource Location (URL)

RESTful queries

Submit request to server via URL
Return result in a structured format
Parse results into a local format

Create a vetiver REST API

R

library(plumber)

pr() |>
  vetiver_api(v) |>
  pr_run()

Python

api = VetiverAPI(v)
api.run()

📝 Create a model API

Instructions

Create a vetiver API for your model and run it locally.

Explore the visual documentation.

How many endpoints are there?

Discuss what you notice with your neighbor.

05:00

What does “deploy” mean?

Where can vetiver deploy?

Enterprise software platforms such as Posit Connect and AWS SageMaker
A public or private cloud, using Docker

How do you make a request of your new API?

R

library(httr2)
url <- "https://info4940.infosci.cornell.edu/tompkins-housing/metadata"

request(url) |>
  req_perform() |>
  resp_body_json()

Python

import requests

url = "https://pub.demo.posit.team/public/seattle-housing-python/metadata" 
print(requests.get(url).content)

How do you make a request of your new API?

Python or R packages like requests or {httr2}
curl
There is special support in vetiver for the /predict endpoint

Any tool that can make an HTTP request can be used to interact with your model API!

Create a vetiver endpoint

You can treat your model API much like it is a local model in memory!

R

library(vetiver)

url <- "https://info4940.infosci.cornell.edu/tompkins-housing/predict"
endpoint <- vetiver_endpoint(url)
predict(endpoint, slice_sample(housing_test, n = 5))

Python

from vetiver.server import predict, vetiver_endpoint

url = "https://info4940.infosci.cornell.edu/tompkins-housing/predict"
endpoint = vetiver_endpoint(url)
predict(endpoint = endpoint, data = X_test.head(5))

📝 Test your predict endpoint

Instructions

Create a vetiver endpoint object for your API.

Predict with your endpoint for new data.

Optional: call another endpoint like /ping or /metadata.

05:00

Wrap-up

Recap

ML models can be deployed as APIs
Use pins to share your models
vetiver can help you bundle, version, and deploy your models

Acknowledgments

Materials derived in part from Intro to MLOps with {vetiver} and licensed under a Creative Commons Attribution 4.0 International (CC BY) License.

Publishing models as APIs

Announcements

Announcements

Learning objectives

Application exercise

ae-14

MLOps

MLOps with {vetiver}

If you develop a model…

If you develop a model…

Tompkins County housing data

Tompkins County housing data

Tompkins County housing data

R

Python

Home prices in Tompkins County, NY

Time for building a model!

Spend your data budget

R

Python

Fit a linear regression model 🚀

R

Python

📝 Fit a model

Create a deployable bundle

Deploy preprocessors and models together

Create a deployable model object

R

Python

Deploy preprocessors and models together

What is wrong with this?

📝 Create a vetiver model object

Version your model

How could you share your resources?

pins 📌

Pin your model

📝 Pin your model

Version your model

Fit a random forest

Version your model

R

Python

📝 Pin a new version

Make it easy to do the right thing

Make it easy to do the right thing

You can deploy your model as a…

Application programming interface (API)

Create a vetiver REST API

R

Python

📝 Create a model API

What does “deploy” mean?

What does “deploy” mean?

Where can vetiver deploy?

How do you make a request of your new API?

R

Python

How do you make a request of your new API?

Create a vetiver endpoint

R

Python

📝 Test your predict endpoint

Wrap-up

Recap

Acknowledgments

`ae-14`