Capabilities
The unit that matters is not the credit — it is the line you are willing to put in the skills section of a CV and then defend to someone who sets the questions. Each of these says what it would take to earn it, and what is still missing.
0 of 12 earned. Most read mostly unmet, which is the system working rather than failing — a bar you can clear on the day you set it was never a bar.
LLM application engineering
1/5 requirements
I can take a model API to a production feature — structured output that cannot be malformed, tool interfaces a model can actually use, a context budget I chose on purpose, and a hallucination guard that fails closed.
5 skills · 0/20 assessed items done
LLM evaluation
1/5 requirements
I can say with a number whether an LLM system is any good, defend the method that produced the number, and show where the measurement itself breaks down.
7 skills · 0/45 assessed items done
Retrieval-augmented generation (RAG)
0/5 requirements
I can design a retrieval system, measure retrieval separately from generation, and name the corpus properties that decide whether embeddings, lexical search or hybrid wins.
7 skills · 0/24 assessed items done
LLM agents and tool use
0/5 requirements
I can choose between a workflow and an agent and justify it, sandbox what a model is allowed to touch, and measure multi-step reliability rather than assert it.
6 skills · 0/21 assessed items done
LLM inference cost and latency optimisation
0/5 requirements
I can account for token spend per run, cut time-to-first-token against a stated budget, and say which optimisations were not worth their complexity.
9 skills · 0/41 assessed items done
Speech interfaces and voice agents
0/5 requirements
I can close a speech loop with a latency budget that survives a real conversation, and handle the case everyone skips — a user interrupting.
5 skills · 0/22 assessed items done
Multimodal and document understanding
0/5 requirements
I can extract structure from real documents and say where a vision-language model is worse than the boring parser it replaced.
2 skills · 0/5 assessed items done
LLM safety and red-teaming
0/5 requirements
I can red-team a system I built, handle personal data deliberately rather than incidentally, and show the attacks that still work.
3 skills · 0/7 assessed items done
AI architecture and build-versus-buy
0/5 requirements
I can decide whether a problem needs an LLM at all, price the alternative, and show a case where the cheaper approach won.
3 skills · 0/29 assessed items done
Transformer architecture
0/5 requirements
I can implement attention and a transformer block from scratch and explain what each part is for without reciting a diagram.
3 skills · 0/17 assessed items done
Fine-tuning and post-training
0/5 requirements
I can curate training data, run a parameter-efficient fine-tune, and show the case where prompting would have been enough.
4 skills · 0/17 assessed items done
Reinforcement learning and RLHF
0/5 requirements
I can explain how preference data becomes a policy, and where reward modelling goes wrong.
4 skills · 0/16 assessed items done
Supporting
Load-bearing under the claims above, and deliberately not claims themselves — nobody is hired for a linear-algebra line, and this programme treats Python as the medium of instruction rather than an achievement.
Engineering foundations
Medium of instruction. Python and deployment are how the work gets done, never what is being claimed — putting them forward as the achievement would misrepresent the programme.
3 skills
Mathematics for machine learning
Load-bearing under deep-learning and fine-tuning, but nobody hires an AI engineer for a linear-algebra line. It earns its place by making other claims defensible.
4 skills
Machine learning foundations
Substrate for evaluation and architecture. The claim it supports is 'I knew a classifier would beat the LLM here', which belongs to architecture.
4 skills
Deep learning
Substrate for transformers and fine-tuning. Claimed through what it enables, not on its own.
4 skills
Text representation
Tokenisation and embeddings are prerequisites for retrieval and transformers. They are a component of those claims rather than a claim of their own.
2 skills