01 Which part of a production AI system runs a trained model on live data?

OrchestrationTrainingIngestionServing and inference

02 What does point-in-time correctness mean?

Every feature uses only data that was available at the moment it would have been usedData is stored in UTCBars are one minute longModels are retrained daily

03 What is training-serving skew?

A slow training jobA biased datasetA GPU shortageA model trained on one feature calculation and served with a different one

04 Why do engineers track p99 latency, not just the average?

It is easier to computeRegulators require itIt is always lower than the averageIt describes the slow tail, which tends to appear when activity spikes

05 What is quantisation?

Splitting data into quantilesStoring model weights at lower numerical precision to speed up inferenceRounding prices to ticksCounting trades

06 What is concept drift?

A new model versionA change in the relationship between inputs and outcomes that the model learnedA change in the input data formatA server moving region

07 In a shadow deployment, a new model...

Only runs at nightReplaces the old model immediatelyRuns on live data and records decisions without acting on themIs trained in secret

08 What is a kill switch?

A data vendor outageA type of stop orderA tested action that stops all automated activity immediatelyA model that predicts crashes

09 Why should position limits be set independently of the model?

To improve latencyExchanges forbid model limitsSo the model cannot override the limits a human has chosenThey are cheaper

10 What does an audit trail record?

Server temperaturesMarketing dataInputs, model version, output and action for every decisionOnly profitable trades

Quiz: AI Infrastructure for Trading Systems