Support Ticket Routing Classifier

A routing model for support tickets that picks the right team and a confidence level. Two model types were compared before choosing, and low-confidence tickets stay in the shared queue.

Machine learning that sends each support ticket to the right team, and leaves unsure ones for a person.

Support tickets landed in one shared inbox and were sorted by whoever picked them up. Misrouted tickets bounced between teams before work started.

Tickets landed in one shared inbox and were sorted by whoever picked them up. Misrouted tickets bounced between teams before anyone started on them.

• •Route each ticket to the right team automatically.
• •Compare model families fairly before choosing one.
• •Never route confidently when the model is unsure.

I built and evaluated the TensorFlow neural network and the XGBoost model, ran the bias and variance diagnostics, and shipped the chosen model behind a Node.js REST endpoint with Redis caching.

Trained a softmax neural network in TensorFlow on ticket text and an XGBoost model on structured fields, compared them on the same held-out set, and deployed the chosen model behind a Node.js endpoint that returns a team and a confidence. Low-confidence tickets stay in the shared queue.

1. 01

   Text and structured features

   Ticket text is vectorized for the neural network; structured fields such as product area and channel feed the tree model.

2. 02

   Neural network (TensorFlow)

   Dense layers with ReLU, L2 regularization and dropout, and a softmax output over teams.

3. 03

   Gradient-boosted trees (XGBoost)

   A tree ensemble trained on the same split for a direct comparison.

4. 04

   Node.js API

   Returns a team and a confidence; below the threshold the ticket stays in the shared queue. Recent results are cached in Redis.

• •Applies neural networks, softmax multiclass output, and decision-tree ensembles from Advanced Learning Algorithms.
• •Uses bias and variance diagnostics and learning curves from the Machine Learning Specialization.
• •A clear abstain path to the shared queue.

model = tf.keras.Sequential([
    tf.keras.layers.Dense(128, activation="relu", kernel_regularizer=tf.keras.regularizers.l2(1e-4)),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(64, activation="relu"),
    tf.keras.layers.Dense(num_teams, activation="linear"),
])
model.compile(optimizer="adam",
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True))
model.fit(X_train, y_train, validation_data=(X_cv, y_cv), class_weight=weights, epochs=40)

1. 01

   Some teams receive far fewer tickets, so class weights and per-class metrics kept the small teams from being ignored.

   Class weights during training and per-class precision and recall during evaluation.

2. 02

   Training and validation curves were used to diagnose high bias versus high variance before adding data or capacity.

   Training and cross-validation curves read before each change, so more data or more capacity was added only when the diagnosis called for it.

3. 03

   A wrong confident route is worse than no route, so the API abstains below a confidence threshold.

   A confidence threshold on the softmax output; tickets below it are not routed.

Security

Ticket text is processed in the service and not written to logs; only the route and confidence are stored.

Speed & growth

One model call per ticket behind a lightweight Node.js endpoint, with Redis caching repeat lookups.

Held-out evaluation per team, threshold tuning on a validation split, and regression tests on a fixed ticket set before each model update.

These are design targets and approach choices for the system, not reported production figures.