import tensorflow as tf
from tensorflow import keras

# Classify 28x28 greyscale images into 10 classes (e.g. product types)
model = keras.Sequential([
    keras.Input(shape=(28, 28)),
    keras.layers.Flatten(),
    keras.layers.Dense(128, activation="relu"),
    keras.layers.Dense(10, activation="softmax"),
])
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
(x_tr, y_tr), (x_te, y_te) = keras.datasets.fashion_mnist.load_data()
model.fit(x_tr / 255.0, y_tr, epochs=3, verbose=0)
print("test accuracy:", round(model.evaluate(x_te / 255.0, y_te, verbose=0)[1], 3))

Using a pre-trained model (transfer learning)
1. Start from a model already trained on a large public dataset
2. Replace its final layer with one for your classes
3. Train only the new layer on your smaller labelled dataset
4. Optionally fine-tune more layers later
Benefit: good results with far less data and compute than training from scratch.

# Simple text classifier for routing support tickets (Keras)
import tensorflow as tf
from tensorflow import keras

texts = tf.constant(["refund not received", "cannot log in", "wrong item delivered", "password reset"])
labels = tf.constant([0, 1, 0, 1])   # 0 = orders team, 1 = account team

vectorize = keras.layers.TextVectorization(max_tokens=1000, output_sequence_length=10)
vectorize.adapt(texts)
model = keras.Sequential([
    keras.Input(shape=(1,), dtype=tf.string),
    vectorize,
    keras.layers.Embedding(1000, 16),
    keras.layers.GlobalAveragePooling1D(),
    keras.layers.Dense(1, activation="sigmoid"),
])
model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy"])
model.fit(tf.reshape(texts, (-1, 1)), labels, epochs=30, verbose=0)
# Real systems need thousands of labelled tickets and proper evaluation

Deep learning use-case canvas
Problem: ______________ (e.g. invoice entry takes too long)
Current cost: time, errors, delays (measured, not guessed)
Data available: type ______ volume ______ labelled? ______
Decision the model supports: ______________
Human in the loop: who checks low-confidence cases?
Success metric: ______________
Risk if wrong: low / medium / high

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
from sklearn.ensemble import HistGradientBoostingClassifier

X, y = load_breast_cancer(return_X_y=True)
baseline = HistGradientBoostingClassifier(random_state=42)
print("baseline accuracy:", cross_val_score(baseline, X, y, cv=5).mean().round(3))
# For tabular data, a strong tree-based baseline is often hard to beat

Responsible deep learning checklist
☐ Data collected with proper consent and rights
☐ Bias checked across relevant groups
☐ Performance measured on realistic test data
☐ Human review for high-impact or low-confidence cases
☐ Monitoring for drift after launch
☐ Clear owner and rollback plan
☐ Explanation available to affected users where needed

Deep learning pilot plan
Problem and metric: ______________
Data: source ______ volume ______ labels ______
Baseline: current process or simple model result ______
Approach: pre-trained model + fine-tuning
Pilot scope: one team, one process, limited time
Human review: ______________
Go/no-go criteria: beats baseline by ____ on agreed metric

Cost checklist
☐ Data labelling effort
☐ Compute for training and serving
☐ Engineering time for integration
☐ Monitoring and retraining
☐ Staff training and process change
Compare with expected benefit before scaling.

Team skills map
Data engineering: pipelines, storage, labelling workflows
ML engineering: models, training, deployment, monitoring
Domain experts: define labels, review outputs
Product/business: metrics, adoption, change management
Compliance: privacy, fairness, documentation