Meridian's engineering team needs to detect an abnormal vibration signature from an aircraft engine within seconds so ground crew can be alerted before the next flight, rather than discovering the anomaly in a report the next morning. Which Big Data approach should the PM push the team toward to enhance this AI capability?

Applying Big Data approaches to enhance AI capability means matching the data architecture to the latency the business need actually requires. Meridian's requirement — seconds-level anomaly alerts — demands a stream-processing architecture that analyzes sensor telemetry in near real time as it arrives, where telemetry is analyzed as it arrives rather than aggregated later. The nightly batch job that aggregates the previous day's sensor readings into a summary report directly fails the stated business requirement: a report the next morning is too late to prevent the next flight's risk, illustrating a common AI-project failure mode of solving the wrong latency problem. The manual quarterly audit of printed maintenance logs by the engineering team ignores the value of the sensor data entirely and reverts to a manual process, defeating the purpose of the initiative. The one-time historical data export used only to train the model, with no ongoing data flow, would support only initial model training, not the ongoing production need for continuous detection — it conflates training-time and inference-time data needs, a distinction this domain tests directly under 'separate training and inference data pipelines.' The exam-relevant judgment: a PM must translate a business latency requirement into the right Big Data processing pattern, streaming versus batch, before the team commits to an architecture.