13 · Personal & Academic Project
Fashion-MNIST Classifier API
Turning a computer-vision model into a deployable prediction service
Business value
The project moved from model experimentation to a reusable microservice that could be called by another application.
01 · Business problem
A high-performing image classifier is not useful to others when it only runs inside a notebook. The challenge was to make inference reliable, repeatable, and accessible through an API.
02 · My approach
I separated training from serving, validated the model on unseen images, and designed the deployment around a lightweight prediction endpoint.
03 · What I built
A ResNet18 classifier trained on 60,000 Fashion-MNIST images, a FastAPI top-k inference endpoint, Gunicorn serving, Docker and Docker Compose packaging, and GCP deployment patterns.
What changed my thinking
The corrections, tradeoffsand decisions behind it.
- 01
Accuracy was only the first milestone
The larger learning came from packaging the model so another system could call it consistently.
- 02
Training and inference needed different workflows
The notebook optimized learning, while the API needed fast loading, validation, and predictable output.
- 03
Deployment exposed hidden assumptions
File formats, preprocessing steps, label order, and model version all had to be preserved outside the training environment.