Cost optimization strategies in Agentic AI - ML Experiment: Train & Evaluate

Experiment - Cost optimization strategies

Problem:You have trained an agentic AI model that performs well but is very expensive to run due to large model size and high inference time.

Current Metrics:Training cost: $500, Inference latency: 1200 ms, Accuracy: 92%

Issue:The model is too costly to deploy in real-time applications because of high inference latency and expensive resource usage.

Your Task

Reduce inference latency below 500 ms and cut deployment cost by at least 50% while keeping accuracy above 88%.

You cannot reduce the training dataset size.

You must keep the model architecture fundamentally the same (no changing to a completely different model).

You can adjust hyperparameters, apply model compression, or optimize inference.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

Agentic AI

import torch
import torch.nn as nn
import torch.quantization

# Assume model is a pretrained PyTorch model
model = ...  # pretrained agentic AI model

# Step 1: Apply pruning
from torch.nn.utils import prune
for name, module in model.named_modules():
    if isinstance(module, nn.Linear):
        prune.l1_unstructured(module, name='weight', amount=0.3)  # prune 30% weights

# Step 2: Convert model to quantized version
model.eval()
model.qconfig = torch.quantization.get_default_qconfig('fbgemm')
torch.quantization.prepare(model, inplace=True)
# Calibration with sample data (dummy example)
input_data = torch.randn(1, 3, 224, 224)
model(input_data)
torch.quantization.convert(model, inplace=True)

# Step 3: Measure inference latency
import time
start = time.time()
_ = model(input_data)
end = time.time()
latency_ms = (end - start) * 1000

print(f'Inference latency after optimization: {latency_ms:.2f} ms')

# Step 4: Evaluate accuracy on validation set (dummy example)
# val_accuracy = evaluate(model, val_loader)  # Assume evaluate function exists
val_accuracy = 89.5  # example after optimization

# Step 5: Estimate cost savings
original_cost = 500
new_cost = original_cost * 0.45  # estimated 55% cost reduction

print(f'Validation accuracy: {val_accuracy}%')
print(f'Estimated deployment cost: ${new_cost}')

Applied 30% pruning on linear layers to reduce model size.

Converted model to 8-bit quantized version to speed up inference.

Measured inference latency showing reduction from 1200 ms to under 500 ms.

Estimated deployment cost reduced by 55% due to smaller model and faster inference.

Results Interpretation

Before Optimization: Inference latency = 1200 ms, Accuracy = 92%, Deployment cost = $500

After Optimization: Inference latency = 480 ms, Accuracy = 89.5%, Deployment cost = $225

Pruning and quantization can significantly reduce model size and inference time, lowering deployment costs while maintaining acceptable accuracy.

Bonus Experiment

Try knowledge distillation to train a smaller student model that mimics the original large model and compare cost and accuracy.

💡 Hint

Use the original model's predictions as soft labels to train a smaller model with fewer parameters.

Practice

1. What is the main goal of cost optimization in agentic AI projects?

easy

A. To increase training time for better accuracy

B. To make AI models as complex as possible

C. To reduce money and resource use while keeping good AI results

D. To use only the newest hardware regardless of cost

Solution

1. Step 1: Understand cost optimization meaning

   Cost optimization means using fewer resources and less money while maintaining good AI performance.

2. Step 2: Match goal with options

   To reduce money and resource use while keeping good AI results clearly states reducing money and resource use while keeping good results, which matches the definition.

3. Final Answer:

   To reduce money and resource use while keeping good AI results -> Option C

4. Quick Check:

   Cost optimization = reduce cost and keep quality [OK]

Hint: Cost optimization means saving money and resources [OK]

Common Mistakes:

• Thinking cost optimization means making models more complex
• Assuming longer training always improves cost
• Ignoring resource use in cost calculations

2. Which of the following is a correct Python syntax to stop training early when validation loss stops improving?

easy

A. early_stopping = EarlyStopping(monitor='val_loss', patience=3)

B. early_stopping = EarlyStopping('val_loss', patience=3)

C. early_stopping = EarlyStopping(monitor=val_loss, patience=3)

D. early_stopping = EarlyStopping(monitor='val_loss' patience=3)

Solution

1. Step 1: Check correct argument syntax for EarlyStopping

   The argument 'monitor' must be a string with the metric name, so monitor='val_loss' is correct.

2. Step 2: Identify correct option

   early_stopping = EarlyStopping(monitor='val_loss', patience=3) uses monitor='val_loss' and patience=3 correctly with commas and quotes.

3. Final Answer:

   early_stopping = EarlyStopping(monitor='val_loss', patience=3) -> Option A

4. Quick Check:

   Correct syntax uses monitor='val_loss' with commas [OK]

Hint: Use quotes around metric names and commas between arguments [OK]

Common Mistakes:

• Missing quotes around 'val_loss'
• Omitting commas between arguments
• Passing variable instead of string for monitor

3. Given this code snippet for training with early stopping, what will be printed?

from tensorflow.keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor='val_loss', patience=2)
history = model.fit(X_train, y_train, epochs=10, validation_split=0.2, callbacks=[early_stopping])
print(len(history.history['loss']))

medium

A. Less than 10

B. 10

C. More than 10

D. Error because EarlyStopping is not defined

Solution

1. Step 1: Understand EarlyStopping behavior

   EarlyStopping stops training early if validation loss does not improve for 'patience' epochs, so training may stop before 10 epochs.

2. Step 2: Predict length of loss history

   Since training can stop early, the number of loss entries will be less than 10.

3. Final Answer:

   Less than 10 -> Option A

4. Quick Check:

   EarlyStopping stops early, so epochs run < 10 [OK]

Hint: EarlyStopping can reduce epochs, so history length is less than max [OK]

Common Mistakes:

• Assuming training always runs full 10 epochs
• Confusing patience with number of epochs
• Thinking EarlyStopping causes errors if used correctly

4. Identify the error in this cost optimization code snippet:

early_stop = EarlyStopping(monitor='val_loss' patience=5)
model.fit(X, y, epochs=20, callbacks=[early_stop])

medium

A. Wrong callback name, should be EarlyStop

B. Missing comma between arguments in EarlyStopping

C. Epochs value too high for cost optimization

D. Callbacks list should be a dictionary, not a list

Solution

1. Step 1: Check EarlyStopping argument syntax

   Arguments must be separated by commas; here, monitor='val_loss' and patience=5 lack a comma.

2. Step 2: Verify callback usage

   EarlyStopping is correct callback name and callbacks parameter expects a list, so no error there.

3. Final Answer:

   Missing comma between arguments in EarlyStopping -> Option B

4. Quick Check:

   Arguments need commas between them [OK]

Hint: Check commas between function arguments carefully [OK]

Common Mistakes:

• Forgetting commas between parameters
• Misnaming callbacks
• Using wrong data type for callbacks argument

5. You want to reduce training cost by using a pre-trained model and early stopping. Which strategy best combines these to optimize cost effectively?

hard

A. Start training a large model from scratch and use early stopping after 20 epochs

B. Use a pre-trained model but train all layers fully without early stopping

C. Train a small model without early stopping to save setup time

D. Use a pre-trained model and fine-tune only last layers with early stopping monitoring validation loss

Solution

1. Step 1: Understand pre-trained model benefits

   Pre-trained models save cost by reusing learned features, reducing training time.

2. Step 2: Combine with early stopping

   Fine-tuning only last layers and using early stopping on validation loss stops training when no improvement, saving resources.

3. Final Answer:

   Use a pre-trained model and fine-tune only last layers with early stopping monitoring validation loss -> Option D

4. Quick Check:

   Pre-trained + early stopping = cost efficient training [OK]

Hint: Fine-tune last layers + early stopping saves cost best [OK]

Common Mistakes:

• Training large models from scratch wastes resources
• Skipping early stopping loses cost savings
• Training all layers fully increases cost unnecessarily