15. Production NLP and MLOps
Understand what changes when an NLP model becomes a service.
01 Learning objectives
By the end of this module, you should be able to explain the concept in plain language, connect it to a use case, and recognize when it is — and is not — appropriate.
-
Objective 1
Think in terms of versioned pipelines, artifacts and monitoring.
-
Objective 2
Understand online vs batch inference.
-
Objective 3
Recognize drift and operational failure modes.
02 Production artifact
A production NLP system is more than a model file. It may include:
- preprocessing/tokenizer version,
- vocabulary or embedding model,
- classifier/model weights,
- thresholds,
- label mapping,
- feature schema,
- runtime/container,
- monitoring rules.
03 Monitoring
Monitor:
- input drift,
- label distribution,
- confidence distribution,
- latency,
- errors,
- business outcome metrics,
- retrieval quality for RAG,
- hallucination/groundedness samples.
04 Batch vs real-time
Batch: large periodic workloads, offline classification, document processing.
Real-time: interactive chatbot, transaction decisioning, autocomplete.
Architecture follows latency, throughput, cost and freshness requirements.
05 Pitfalls and nuances
These are the mistakes most likely to create a wrong mental model or a weak implementation.
-
Pitfall 1
Updating a tokenizer without updating the model.
-
Pitfall 2
Threshold changes without governance/testing.
-
Pitfall 3
Monitoring infrastructure metrics but not semantic quality.
06 Checkpoint questions
Try answering these without looking back at the module. If the explanation is clear in your own words, you are ready to continue.
-
Check 1
Why must tokenizer and model versions travel together?
-
Check 2
What could semantic drift look like?