Intermediate

AI Solutions (25%)

Domain 3 of the CompTIA AI+ exam — AI implementation strategies, integrating AI with existing systems, deployment options, monitoring, and real-world applications.

AI Deployment Strategies

Deployment Options

  • Cloud deployment — Run AI models on cloud infrastructure (AWS, Azure, GCP). Benefits: scalability, GPU access, managed services. Risks: cost, latency, data privacy.
  • On-premises deployment — Run models on local servers. Benefits: data control, low latency, compliance. Risks: hardware costs, maintenance.
  • Edge deployment — Run models on edge devices (phones, IoT, cameras). Benefits: low latency, works offline, reduces bandwidth. Risks: limited compute, model size constraints.
  • Hybrid deployment — Combine cloud and edge. Heavy processing in the cloud, lightweight inference at the edge.

Model Serving Patterns

  • Real-time inference — Model responds immediately to each request (REST API, gRPC). Use for: chatbots, recommendations, fraud detection.
  • Batch inference — Process large datasets in bulk on a schedule. Use for: report generation, data enrichment, nightly predictions.
  • Streaming inference — Process continuous data streams in real-time. Use for: sensor data, log analysis, real-time monitoring.

Integration with Existing Systems

  • API integration — Expose AI models as REST or GraphQL APIs that other applications consume
  • Embedded AI — Build AI directly into existing applications (adding smart features to an existing product)
  • Microservices — Deploy AI as independent services in a microservices architecture
  • ETL pipelines — Integrate AI into data processing pipelines for transformation and enrichment
💡
Key concept — MLOps: MLOps (Machine Learning Operations) applies DevOps practices to ML systems. It includes automated training pipelines, CI/CD for models, monitoring, versioning, and reproducibility. MLOps is how organizations scale AI from experiments to production.

Model Monitoring

  • Data drift — Input data distribution changes over time (e.g., customer behavior shifts). Requires retraining.
  • Concept drift — The relationship between inputs and outputs changes (e.g., what defines "spam" evolves).
  • Model degradation — Performance metrics (accuracy, latency) worsen over time.
  • Performance metrics — Track accuracy, latency, throughput, error rates in production.
  • A/B testing — Compare new model version against the current one with real traffic to validate improvements.

Real-World AI Applications

Healthcare

  • Medical image analysis (X-ray, MRI, CT scan interpretation)
  • Drug discovery and molecular modeling
  • Patient risk prediction and treatment recommendations

Finance

  • Fraud detection (real-time transaction monitoring)
  • Algorithmic trading and market prediction
  • Credit scoring and risk assessment

Manufacturing

  • Predictive maintenance (predict equipment failures before they happen)
  • Quality inspection (visual defect detection on production lines)
  • Supply chain optimization

Retail

  • Recommendation engines (product suggestions)
  • Demand forecasting and inventory management
  • Customer sentiment analysis from reviews

AI Infrastructure

  • GPUs — Essential for training deep learning models. NVIDIA is the dominant provider.
  • TPUs — Google's custom AI accelerators, optimized for TensorFlow workloads.
  • Cloud AI services — Pre-built AI APIs (vision, speech, language) that require no ML expertise.
  • Containers — Package AI models with their dependencies for consistent deployment (Docker, Kubernetes).
  • Model registries — Version and manage trained models (MLflow, AWS SageMaker Model Registry).

Practice Questions

📝
Q1: A factory needs to run a defect detection model on cameras directly on the production floor with no cloud connectivity. Which deployment strategy should they use?

A) Cloud deployment
B) Edge deployment
C) Batch inference
D) Hybrid deployment
Show Answer

B) Edge deployment. Running the model directly on cameras or local devices at the production floor (edge) eliminates the need for cloud connectivity. Edge deployment provides low latency (critical for real-time defect detection) and works offline. The model must be small enough to run on edge hardware.

📝
Q2: A model that was 95% accurate six months ago is now only 85% accurate, even though the model has not changed. What is the most likely cause?

A) Overfitting
B) Data drift
C) Underfitting
D) Class imbalance
Show Answer

B) Data drift. When the model has not changed but performance degrades over time, the most likely cause is data drift — the input data distribution has changed since the model was trained. The solution is to retrain the model with recent data that reflects current patterns.

📝
Q3: A company wants to compare a new recommendation model against the current one using real customer traffic. Which technique should they use?

A) Cross-validation
B) A/B testing
C) Data augmentation
D) Hyperparameter tuning
Show Answer

B) A/B testing. A/B testing splits real traffic between the current model (control) and the new model (treatment) to compare performance with actual users. Cross-validation is a training technique, not a production comparison method.

📝
Q4: Which deployment pattern processes thousands of customer records overnight to generate predictions for a morning report?

A) Real-time inference
B) Streaming inference
C) Batch inference
D) Edge inference
Show Answer

C) Batch inference. Batch inference processes large volumes of data in bulk on a schedule (e.g., nightly). It is cost-effective for non-time-sensitive predictions like generating reports, scoring customer databases, or enriching datasets.

📝
Q5: Which practice applies DevOps principles to machine learning systems, including automated training, CI/CD for models, and monitoring?

A) DataOps
B) MLOps
C) AIOps
D) DevSecOps
Show Answer

B) MLOps. MLOps (Machine Learning Operations) combines ML, DevOps, and data engineering practices to automate and manage the ML lifecycle in production. It covers automated training pipelines, model versioning, CI/CD, monitoring, and reproducibility.