Stop 12 · Model Deployment & MLOpsTopics · 2/8

Building an API for a Model (Flask / FastAPI)

In short

A website or app can't import your notebook. Wrap the model in a small web API and anything that can send an HTTP request, from a phone app to a spreadsheet script, can get predictions.

Wrapping a trained model in a lightweight web API so other applications can send it data and get predictions back — FastAPI is the modern, faster-growing choice with built-in data validation.

The mascot at a takeaway window takes an order slip and hands back a food parcel.

Think of it like this

The takeaway counter

At a takeaway window, you hand over an order slip and get a parcel back. You never enter the kitchen or meet the cook. The counter checks your slip is valid and passes it inside. An API is that counter, and your model is the kitchen.

Watch it happen

One prediction request

Client app

FastAPI

Model

POST /predict

1/5The app sends JSON with four flower measurements.

The big ideas

What you need to know

FastAPI

Type hints become input validation and automatic docs at /docs. A popular modern choice for ML APIs.

Flask

Older, minimal and very common. You handle validation yourself or add a library.

Load once

Load the model when the server starts, not inside the request function, or every call gets slow.

Validate inputs

Real users send missing fields, text in number fields and extreme values. Reject them clearly.

See it in code

Walk through the code

python

1import joblib2from fastapi import FastAPI3from pydantic import BaseModel4 5app = FastAPI()6model = joblib.load("iris_model.joblib")          # load once at startup7 8class Flower(BaseModel):9    sepal_length: float10    sepal_width: float11    petal_length: float12    petal_width: float13@app.post("/predict")14def predict(f: Flower):15    pred = model.predict([[f.sepal_length, f.sepal_width, f.petal_length, f.petal_width]])16    return {"class_id": int(pred[0])}

1/4Create the app and load the saved model a single time.

Save as main.py, install fastapi and uvicorn, then run uvicorn main:app --reload and open /docs to try it.

Save yourself the pain

Common mistakes

Oops

Loading the model file inside the predict function.

Do this instead

Load it once at startup. Reading from disk on every request wastes time.

Oops

Returning NumPy types directly in the response.

Do this instead

Convert to int, float or lists first. Some NumPy types can't be turned into JSON.

Your mission

Serve your saved model

~30 min

Quick check

Why load the model at server startup instead of inside the request handler?

Go deeper

1Build a simple FastAPI endpoint that loads a model and returns a prediction.
2Test your API locally with a sample request using a tool like curl or Postman.
3Learn to validate incoming request data before passing it to the model.