06
Production MLOps and Certification Prep
Ship and keep models healthy: serving, quantisation, drift, LLM inference and responsible AI, mapped to the 2026 Google Cloud, AWS, Azure and NVIDIA exam blueprints.
- Lessons
- 5
- Study time
- 2.5 h
- Access
- Pro track
Syllabus
- 6.1
Quantisation and efficient inference
INT8, FP8 and INT4 weight formats, the distortion they introduce, and the latency they buy back.
32 minProLive 3D lab and Python challengeintermediate- Quantise and dequantise a weight tensor with per-channel scales
- Measure the accuracy and latency trade-off
- Choose between post-training quantisation and QAT
- 6.2
Distribution shift and drift monitoring
Covariate, label and concept drift, the statistics that detect them, and when to retrain.
28 minProLive 3D lab and Python challengeintermediate- Distinguish covariate, prior and concept shift
- Detect drift with PSI and KS statistics
- Design a retraining trigger
- 6.3
Training pipelines, feature stores and serving
Reproducible pipelines, online/offline feature parity, and batch versus real-time serving architectures.
30 minProLive 3D lab and Python challengeintermediate- Structure a reproducible training pipeline
- Avoid training-serving skew with a feature store
- Pick a serving pattern for a latency budget
- 6.4
LLM serving: KV cache, batching and speculative decoding
Why decoding is memory-bound, how the KV cache grows, and how paged memory, continuous batching and draft models raise throughput.
32 minProLive 3D lab and Python challengeadvanced- Size the KV cache for a model, context length and batch
- Compare static, dynamic and continuous batching
- Explain speculative decoding and when it pays off
- 6.5
Responsible AI, fairness and generative AI security
Group fairness metrics and their trade-offs, explainability, and the prompt-injection and data-exfiltration threats every AI exam now covers.
30 minProLive 3D lab and Python challengeintermediate- Compute demographic parity, equal opportunity and calibration across groups
- Explain why fairness criteria conflict when base rates differ
- Mitigate prompt injection, data exfiltration and unsafe output with layered guardrails