ML·Shipped
Learning path
Three small ML projects that got me from first CNN to a fine-tuned transformer
98.24% / 90.54% / 88.4%
MNIST / Fashion-MNIST / DistilBERT
- Role
- Built all three classifiers and their Streamlit apps as learning projects.
- Team
- UNKNOWN
- When
- 6-7 Jul 2026
Problem
Learn the fundamentals of deep learning by building and deploying three small classifiers.
Approach
An MNIST digit classifier and a Fashion-MNIST classifier, then a DistilBERT sentiment model fine-tuned with HuggingFace. Each one has a Streamlit app.
Result
MNIST: 98.24% accuracy. Fashion-MNIST: 90.54% accuracy. DistilBERT sentiment: 88.4% accuracy.
Features
- MNIST digit classifier (6 Jul)
- Fashion-MNIST classifier (6 Jul)
- DistilBERT sentiment analysis (7 Jul)
- Streamlit app for each one
What I learned
Together they cover CNNs, DataLoaders, training loops, tokenization, the Trainer API and deployment.
Tech
PythonPyTorchCNNsHuggingFaceDistilBERTStreamlit
Screens
MNIST app: a hand-drawn 9 on the drawing board is recognised
1 / 6MNIST app: a hand-drawn 9 on the drawing board is recognised
1 / 6keep scrolling ↓