🎯 Why This Episode Matters

In Episode 2, we trained a model that looked perfect inside the notebook…
but collapsed the second it saw real user messages.

In Episode 3, we take that same broken model and deploy it behind an API using FastAPI + Docker.

Because this is how ML models are deployed in the real world.
And this is exactly where things break in production.

If you’ve ever deployed an ML model and wondered:

“Everything looks healthy… then why are the predictions wrong?”

This episode explains it clearly.

▶️ Watch the Full Episode

🎥 YouTube: Deployed the Model… But It Failed — MLOps EP3
👉 https://youtube.com/@learnwithdevopsengineer

(Full code available for newsletter subscribers — link at the bottom.)

📌 The Reality of ML Deployment — Healthy API, Broken Predictions

This episode shows the most dangerous ML failure:

The model is wrong…
but everything else looks fine.

When we deploy the Episode 2 model behind an API:

• The server runs normally

• Logs look clean

• CPU and memory are stable

• Docker container is healthy

• API returns 200 OK

And yet…

The predictions are incorrect.
Silently.

This is the kind of failure that causes real production incidents.

🧱 How We Built the FastAPI Service

To stay realistic, we create a simple FastAPI inference service:

• Load the saved model

• Load the saved vectorizer

• Accept incoming text

• Vectorize

• Predict

• Return the result

No validation.
No preprocessing.
No monitoring.
No versioning.
No drift detection.

And this is exactly how most first-time ML deployments happen inside companies.

Which is why they fail.

🐳 Dockerizing the Model

After confirming the API works locally, we move to the Docker step.

The Dockerfile is intentionally simple:

• Python base image

• Install FastAPI + scikit-learn

• Copy model + vectorizer

• Start the server

This gives us a clean, lightweight model container — very similar to how teams deploy ML microservices.

And right after deployment…
the broken predictions continue.
Because Docker does not magically fix bad modeling decisions.

🔍 The Live Failure Test

In Swagger UI, we test two groups of messages:

1️⃣ Messages similar to training data

The API behaves well.
It returns correct predictions.

2️⃣ Real-world messy user messages

Suddenly the model produces:

• Wrong categories

• Incorrect labels

• Weak patterns

• Zero understanding of messy text

The deployment is not the problem.
The model itself is not prepared for real-world data.

This is where MLOps becomes necessary.

📊 The Lesson — Why MLOps Exists

This episode demonstrates the exact reason MLOps roles exist:

• Systems can look healthy

• APIs can respond normally

• Infrastructure can be perfect

…but the ML logic inside can still be failing.

ML systems don’t just need deployment.
They need:

• Input validation

• Output monitoring

• Drift detection

• Model versioning

• Proper packaging

• Pipelines

• Reproducibility

• Evaluation gates

These are the pieces we start building from Episode 4 onwards.

🚀 Coming Up in Episode 4

Episode 4 is where the real engineering begins:

• Packaging models properly

• Saving vectorizers the right way

• Adding evaluation checks

• Monitoring inputs and outputs

• Setting up a drift detection mechanism

• Using MLflow Model Registry

• Versioning models

• Improving the deployment strategy

This is the turning point of the entire series.

🔗 Full Video + Code Access

🎥 Watch Episode 3: https://youtube.com/@learnwithdevopsengineer
📬 Code + Labs: https://learnwithdevopsengineer.beehiiv.com/subscribe

Subscribers get:

• Full project code

• FastAPI deployment files

• Docker setup

• Inference scripts

• Drift testing samples

• Interview questions

• Real MLOps exercises

If you want to learn MLOps the real way — this is the starting point.

💼 Need MLOps/DevOps Help?

If you’re working on:

• CI/CD pipelines

• Docker / Jenkins

• MLflow tracking

• FastAPI deployments

• Infrastructure automation

• Monitoring & alerting

• Kubernetes

• Cloud cost optimization

• MLOps workflows

You can consult me directly.

Reply to this email or message me on YouTube/Instagram.

— Arbaz
📺 YouTube: Learn with DevOps Engineer
📬 Newsletter: learnwithdevopsengineer.beehiiv.com/subscribe