Skip to content
LearnMLByDoing
Free
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

  1. 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
  2. 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
  3. 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
  4. 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
  5. 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