AI Applications, Tools, and Deployment
Core concepts
Deployment turns a model into a reliable service. Know the lifecycle: problem framing, data preparation, training, evaluation, deployment, monitoring, and iteration. Understand inference (serving predictions) versus training, and the trade-off between batch (scheduled, high-throughput) and real-time (low-latency) serving. Recognize APIs, model endpoints, and MLOps practices-versioning, CI/CD, automated testing, rollback-that keep systems reproducible.
Operations and risk
Key concepts: model drift (input or concept drift degrades performance over time), data drift, monitoring for accuracy and fairness, latency and cost budgets, and scalability. Know that deployment includes human oversight and fallback paths for high-stakes uses. Understand edge versus cloud deployment: edge reduces latency and can improve privacy but constrains compute; cloud offers scale at the cost of network dependence.
Traps
Questions imply deployment ends the project; in reality monitoring and retraining are ongoing. Watch for confusing drift (world changes) with bugs (code errors)-both degrade output but need different fixes. Another trap: assuming a model that performed well offline will perform identically in production; data distributions differ. Also beware treating a vendor tool as a substitute for evaluation-tools accelerate work but do not guarantee fitness.
Rule of thumb
Before launch, define what success looks like in production (metric, threshold, monitoring cadence, owner, rollback trigger). No monitoring plan means no deployment.
Worked example
A demand-forecasting model is accurate at launch but degrades after a market shift-concept drift. The response is not retraining blindly but detecting drift via monitoring, investigating the cause, then retraining on recent data and revalidating before redeployment.