⚡Fixing the Model — Building the FIRST Real MLOps Pipeline (EP4)

MLOps Series — Preprocessing + Packaging Done the Right Way

Author

Arbaz M
November 27, 2025

🎯 Why This Episode Matters

In Episode 3, we deployed our model behind FastAPI + Docker and watched it fail instantly on real-world inputs — even though the API, container, and server were all healthy.

Episode 4 is where we FIX the root cause.

This is the first episode where we move from “quick ML experiments” to actual MLOps engineering:

  • proper preprocessing

  • proper packaging

  • consistent training + inference

  • versioning

  • reproducible behavior

This episode marks the moment the model becomes production-ready.

📌 Why Episode 3 Failed

This is the most common real-world ML failure.

Our model wasn’t wrong.
Our preprocessing was wrong.

Here’s what caused the failure:

  • No preprocessing pipeline

  • Vectorizer not saved or versioned correctly

  • Input cleaning mismatch

  • Tokenization mismatch

  • Model and vectorizer stored separately

  • Training logic ≠ inference logic

This is EXACTLY what happens in companies:

ML engineer saves model.pkl
Forgets the vectorizer
DevOps deploys it
Vectorizer mismatch
Model fails silently
End users get nonsense predictions

Episode 4 is the fix for this entire class of failures.

🧱 The Fix — Build a REAL Machine Learning Pipeline

We replace the broken approach with a proper scikit-learn Pipeline: 

Pipeline([
    ("vectorizer", CountVectorizer()),
    ("model", LogisticRegression())
])

One object.
One file.
One save.
One load.

No mismatches.
No separate artifacts.
No confusion between training and serving.

This is how ML models SHOULD be packaged.

🧪 Training With the New Pipeline

We rewrite the training script (train_v2.py) to:

  • use the pipeline

  • train end-to-end preprocessing + model

  • evaluate correctly

  • save a single pipeline.pkl

  • log everything in MLflow

This version is:

  • cleaner

  • reproducible

  • usable in any environment

  • ready for real-world deployment

It is already 10x more production-ready than our Episode 2 model.

📁 Inference With the Correct Pipeline

We also build a new inference script (infer_v2.py) that loads the full pipeline and tests it on real-world messages.

You’ll immediately notice:

  • noisy messages perform better

  • predictions stabilize

  • behavior is more consistent

  • accuracy is more realistic

This is the power of combining preprocessing + model into one pipeline.

🏃 Updating the FastAPI Service

We update the API:

  • load only pipeline.pkl

  • remove manual vectorizer logic

  • remove mismatched transformations

  • simplify inference

  • reduce risk

The new API is cleaner, safer, and exactly how real companies deploy ML today.

🐳 Dockerfile v2 — Production-Friendly

The updated Dockerfile includes the new pipeline and the v2 API.

We build and run: 

docker build -t ml-api-v2 .
docker run -p 8000:8000 ml-api-v2

Same deployment process — but with a far more stable model inside.

🧠 What EP4 Teaches You

Episode 4 shows the core idea of MLOps:

ML models do not fail because of accuracy.
They fail because training ≠ serving.

This episode fixes:

  • vectorizer mismatch

  • preprocessing mismatch

  • dependency mismatch

  • versioning issues

  • inconsistent logic

This is the first time in the series where our model becomes truly production-ready.

🚀 Coming Up in Episode 5

Next episode takes the next big step:

  • automated retraining

  • validation checks

  • CI/CD for model updates

  • batch evaluation

  • data drift strategies

  • MLflow Model Registry

  • proper versioning

  • improved deployment

This is where the pipeline becomes a real system.

🔗 Full Video + Code Access

Subscribers get:

  • full project code

  • all pipelines

  • Dockerfiles

  • model registry examples

  • drift testing scripts

  • real incident simulations

  • interview prep

This entire series will accelerate your MLOps career.

💼 Need DevOps/MLOps Help?

If you’re building:

  • CI/CD pipelines

  • Docker or Jenkins setups

  • MLflow tracking

  • FastAPI deployments

  • production ML pipelines

  • monitoring and alerting

  • Kubernetes workloads

  • cloud cost optimization

You can consult me directly.

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