How TensorFlow Helps You Land a QA Tester Job Faster

# tiny_model.py — create a model to practise testing on
import numpy as np
import tensorflow as tf

rng = np.random.default_rng(42)
x = rng.random((1000, 3)).astype("float32")       # 3 numeric features
y = (x[:, 0] + x[:, 1] > 1.0).astype("int32")      # simple rule as label

model = tf.keras.Sequential([
    tf.keras.layers.Input(shape=(3,)),
    tf.keras.layers.Dense(8, activation="relu"),
    tf.keras.layers.Dense(1, activation="sigmoid"),
])
model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy"])
model.fit(x, y, epochs=20, verbose=0)
model.save("approval_model.keras")

# predict.py — what a tester does with a delivered model
import numpy as np
import tensorflow as tf

model = tf.keras.models.load_model("approval_model.keras")

samples = np.array([
    [0.9, 0.8, 0.1],   # expected: approve (sum > 1)
    [0.1, 0.2, 0.9],   # expected: reject
], dtype="float32")

probs = model.predict(samples, verbose=0).ravel()
print(probs.round(3))
print((probs >= 0.5).astype(int))   # -> [1 0] if the model learned the rule

# test_model_quality.py — run with pytest
import numpy as np
import tensorflow as tf

model = tf.keras.models.load_model("approval_model.keras")
rng = np.random.default_rng(7)
x_test = rng.random((300, 3)).astype("float32")
y_test = (x_test[:, 0] + x_test[:, 1] > 1.0).astype("int32")

def test_accuracy_meets_threshold():
    preds = (model.predict(x_test, verbose=0).ravel() >= 0.5).astype(int)
    accuracy = (preds == y_test).mean()
    assert accuracy >= 0.85, f"accuracy {accuracy:.2f} below 0.85"

# Check accuracy separately for a slice of the data
import numpy as np

def accuracy(model, x, y):
    preds = (model.predict(x, verbose=0).ravel() >= 0.5).astype(int)
    return (preds == y).mean()

def test_accuracy_on_high_feature3_slice():
    mask = x_test[:, 2] > 0.8          # e.g. a specific customer group
    overall = accuracy(model, x_test, y_test)
    slice_acc = accuracy(model, x_test[mask], y_test[mask])
    # The slice should not be much worse than overall
    assert slice_acc >= overall - 0.05, (overall, slice_acc)

# (uses model, x_test, y_test from the threshold test file)

# Small noise should rarely change the decision
import numpy as np

def test_predictions_stable_under_small_noise():
    rng = np.random.default_rng(1)
    # Use points away from the decision boundary (sum far from 1.0)
    x = rng.random((200, 3)).astype("float32")
    x = x[np.abs(x[:, 0] + x[:, 1] - 1.0) > 0.2]
    noisy = x + rng.normal(0, 0.005, x.shape).astype("float32")
    base = model.predict(x, verbose=0).ravel() >= 0.5
    after = model.predict(noisy, verbose=0).ravel() >= 0.5
    flip_rate = (base != after).mean()
    assert flip_rate <= 0.02, f"flip rate {flip_rate:.3f}"

# Validate inputs before they reach the model
import numpy as np

def validate_features(rows):
    arr = np.asarray(rows, dtype="float64")       # raises on non-numeric text
    if arr.ndim != 2 or arr.shape[1] != 3:
        raise ValueError("expected rows of 3 features")
    if np.isnan(arr).any():
        raise ValueError("missing values present")
    if (arr < 0).any() or (arr > 1).any():
        raise ValueError("features must be scaled to 0-1")
    return arr.astype("float32")

# Tester's cases: wrong column count, NaN, 1.5, "abc"

ML FEATURE TEST PLAN

Feature        : ____________________
Model version  : ________  Data version: ________
Agreed metrics : accuracy >= __  recall >= __
Key slices     : city / age band / device / new users

Tests
[ ] Input schema and ranges
[ ] Overall metric thresholds on held-out test data
[ ] Metrics per slice within agreed gap
[ ] Perturbation stability
[ ] Unseen category / out-of-range handling
[ ] API: validation errors, latency, timeouts
[ ] Compare with previous model version

ML DEFECT REPORT

Title    : Approval accuracy much lower for feature3 > 0.8 slice
Model    : approval_model v1.3   Test data: test_2026_10
Metric   : overall 0.91, slice 0.78 (agreed max gap 0.05)
Evidence : test_model_quality.py::test_accuracy_on_high_feature3_slice
Impact   : One customer group gets wrong decisions more often
Suggested: check training data coverage for this slice

# ML Testing Portfolio Project

## What it shows
Testing a TensorFlow classification model like a QA engineer.

## Tests included
- Accuracy threshold on held-out data
- Slice-based accuracy comparison
- Perturbation stability
- Input validation for the prediction API

## How to run
pip install tensorflow numpy pytest
pytest -v

## What I learned
(3-4 lines in your own words)