Expert AI: architecture, training and production systems

About this track

The end of the AI path, for engineers who already build with models and now need to understand and own the whole stack. This track works through the transformer architecture and the attention computation itself, how a frontier model is pretrained and aligned, scaling laws and the economics of compute, fine-tuning and low-rank adapters, inference optimisation, advanced retrieval, the architecture of agent systems that survive contact with production, evaluation with statistical rigour, operations, and the governance and legal obligations that now accompany deployment. It expects comfort with linear algebra notation, probability and Python.

Every lesson is written and hosted here on Yanjye — you never leave the site. Work through them in order, then sit the exam to earn your certificate.

Create a free account or log in to track your progress and earn the certificate.

Lessons

1. 1
   Inside the transformer

   The architecture that every current language model is a variation of.

   ~30 min
2. 2
   Attention, computed step by step

   Queries, keys, values, masking, and the quadratic cost that shapes everything.

   ~30 min
3. 3
   How a frontier model is trained

   Pretraining, supervised fine-tuning, preference optimisation.

   ~28 min
4. 4
   Scaling laws and the economics of compute

   Why bigger worked, what compute-optimal means, and where the limits are.

   ~25 min account needed
5. 5
   Fine-tuning, LoRA, and when not to

   Parameter-efficient adaptation, and the decision that precedes it.

   ~26 min account needed
6. 6
   Alignment, safety and adversarial robustness

   What alignment techniques do, what they do not, and how to test.

   ~26 min account needed
7. 7
   Serving models: the inference stack

   Prefill and decode, batching, quantisation, speculative decoding.

   ~28 min account needed
8. 8
   Advanced retrieval architectures

   Fusion, rerankers, late interaction, graphs, and the long-context question.

   ~26 min account needed
9. 9
   Agent architecture that survives production

   Context engineering, decomposition, multi-agent, and knowing when not to.

   ~26 min account needed
10. 10
    Evaluation with statistical rigour

    Sample size, judge validation, paired tests and online measurement.

    ~26 min account needed
11. 11
    Operating an AI system

    Observability, incident response, model upgrades and regression control.

    ~24 min account needed
12. 12
    Governance, risk and the law

    Regulatory direction, documentation, and the obligations that now attach to deployment.

    ~24 min account needed