AI Learning Ramp
A three-times-a-week course plan to ramp from GenAI query systems into OpenAI or Anthropic style AI engineering interviews. Course 1 is ready now: one hour on LLM serving foundations.
This ramp prioritizes interview-important systems judgment before breadth. The first phase builds the mental model for how LLM requests move through serving infrastructure. The second connects that infra to retrieval, query engines, tools, agents, evals, reliability, security, and cost. The final phase turns the material into system-design answers, tradeoff trees, and mock interview prompts.
One-hour objective: after this session, you should be able to whiteboard the end-to-end path of a production LLM request, explain prefill vs decode, and name the core latency, throughput, and memory bottlenecks.
Anchor on a single user request: prompt assembly, routing, inference, streaming, post-processing, and observability.
Read the concise lifecycle article and capture the difference between TTFT and inter-token latency.
Read the selected continuous-batching sections and connect request length, concurrency, GPU utilization, and memory pressure.
Use the canonical vLLM blog to connect PagedAttention to virtual memory and higher-throughput serving.
Sketch tokenize, prefill, KV cache allocation, decode loop, streaming, finish condition, cache release, metrics, and logs.
Answer: "Why can a long schema prompt hurt time-to-first-token, and why can many short follow-up queries still pressure decode throughput?"
Required reading is now three articles, about 40 minutes total. The fourth item is an optional transformer refresher only if the attention/KV-cache basics feel stale.
A current, practical explanation of prefill, decode, KV cache, TTFT, and inter-token latency. It gives you the vocabulary interviewers expect when discussing LLM serving behavior.
Extract: prefill is the prompt-processing phase; decode is the token-by-token phase that repeatedly reuses cached K/V state.
The best fit for Course 1 because it derives KV caching, chunked prefill, and continuous batching from the user-visible behavior of chat systems. It is more directly useful here than a broad optimization survey.
Extract: continuous batching is iteration-level scheduling; it keeps GPU work full while requests enter and leave at different times.
The canonical vLLM introduction. It explains why KV cache fragmentation limits batching and how PagedAttention applies virtual-memory ideas to LLM serving.
Extract: PagedAttention is a memory-management design; better cache packing lets the serving layer batch more active sequences.
Use only as a quick refresher if Q/K/V, attention, and autoregressive decoding are not fresh. It is background, not a required AI infra article.
Extract: why decoding produces one token at a time even though each forward pass uses highly parallel matrix operations.
The original first-course list was too broad. Redis, Hugging Face, and the vLLM blog are the strongest required set for a one-hour serving-foundations session. Jay Alammar is useful but optional. NVIDIA's inference optimization article is high-quality, but it is better for Course 2 because it introduces too many knobs at once. The Anthropic and OpenAI agent articles move to the agent architecture week.
Use this as the 10-minute close. Speak out loud, as if an OpenAI or Anthropic interviewer asked you to design the system.
Cadence: three sessions per week. The 24-session core runs from now into mid-July, leaving the second half of July for mock interviews, weak-spot repair, and company-specific prep.