Django for AI: Deploying Machine Learning Models with Django

DjangoCon US 2025 talk. Full notes available here.

Transcript

1. Django for AI Will Vincent By Will Vincent DjangoCon US (September 2025)
3. Django Survey 2024
4. Django Survey 2024
6. What is Data Science?
7. FastAPI GitHub stars
8. FastAPI GitHub stars
10. Why Not ? • Slow • Old • Too bloated/hard to learn
11. Web ≠ API Endpoints
13. Part 1: Classic ML Part 2: LLMs Gameplan
14. Classic Machine Learning
15. Artificial Intelligence (AI) Machine Learning (ML)
17. Iris Dataset
18. Jupyter Notebook - Setup
19. Iris Dataset
22. Jupyter code... 1. Imports: pandas & scikit-learn 2. Load iris.csv file 3. Split into training (80%) and testing data (20%) 4. Train the SVM Model 5. Make predictions and evaluate its accuracy 6. Save the model using joblib
23. Jupyter Notebook # Train the model model = SVC(gamma="auto") model.fit(X_train, y_train) # Make predictions on validation dataset predictions = model.predict(X_test) # Evaluate model accuracy accuracy = accuracy_score(y_test, predictions) print(f"Model Accuracy: {accuracy:.2f}")
24. Live Demo?
25. Django + ML
26. Gameplan • Create new Django project • Load joblib file • Forms for users to enter predictions -> see results • Store user info in the database • Deployment?
27. ├── django_project │ ├── __init__.py │ ├── asgi.py │ ├── settings.py │ ├── urls.py │ └── wsgi.py ├── iris.joblib # hi there! ├── manage.py ├── predict │ ├── __init__.py │ ├── admin.py │ ├── apps.py │ ├── migrations │ │ └── __init__.py │ ├── models.py │ ├── tests.py │ └── views.py └── templates
28. # predict/views.py import joblib import numpy as np from django.shortcuts import render # Load the trained model model = joblib.load("iris.joblib") def predict(request): prediction = None if request.method == "POST": try: # Get input values from the form sepal_length = float(request.POST.get("sepal_length")) sepal_width = float(request.POST.get("sepal_width")) petal_length = float(request.POST.get("petal_length")) petal_width = float(request.POST.get("petal_width")) # Make a prediction features = np.array( [[sepal_length, sepal_width, petal_length, petal_width]] ) prediction = model.predict(features)[0] except Exception as e: prediction = f"Error: {str(e)}" return render(request, "predict.html", {"prediction": prediction})
29. Add a Model to Store Predictions # predict/models.py from django.db import models class IrisPrediction(models.Model): sepal_length = models.FloatField() sepal_width = models.FloatField() petal_length = models.FloatField() petal_width = models.FloatField() prediction = models.CharField(max_length=100) created_at = models.DateTimeField(auto_now_add=True) def __str__(self): return f"{self.prediction} ({self.created_at.strftime('%Y-%m-%d %H:%M:%S')})"
30. DjangoForDataScience.com
32. Update Admin
33. Deployment...
34. Deployment Checklist • configure static files and install WhiteNoise • add environment variables with environs • create a .env file and update the .gitignore file • update DEBUG, ALLOWED_HOSTS, SECRET_KEY, and CSRF_TRUSTED_ORIGINS • update DATABASES to run PostgreSQL in production and install psycopg • install Gunicorn as a production WSGI server • create a Procfile • update the requirements.txt file • create a new Heroku project, push the code to it, and start a dyno web process
35. Where to Go from Here? 1. Authentication Security APIs
36. LLMs Under-the-Hood
37. Artificial Intelligence (AI) Machine Learning (ML) Neural Networks/Deep Learning Large Language Models (LLMs)
38. Neural Network
39. How do LLMs Work?
40. Neural Network
50. What is Inference? Model OUTPUT INPUT Text Image Soundwave Molecules Videos Text Image Soundwave Molecules Videos Source: DataCamp Video -> Understanding LLM Inference | NVIDIA Experts Deconstruct How AI Works
51. LLM Inference? GPU(s) OUTPUT/ PROMPUT PROMPT/INPUT Why is Django the best web framework? Django is a great choice for ... Every output token generated gets fed in as a prompt
52. LLM Inference - Processing GPU(s) OUTPUT/ PROMPUT PROMPT/INPUT Why is Django the best web framework? Django is a great choice for ... Tokenization De-Tokenization Initial Prompt Processing Output Prompt Processing
59. Questions? @wsvincent.bsky.social @wsvincent -- djangoforai, django_irisml, iris_ml fosstodon.org/@wsvincent @william-s-vincent