Every Applied AI topic, ordered easy → hard.

Tokens, attention, context windows, sampling, prompting vs RAG vs fine-tuning, hallucinations, evals, cost and latency: the conceptual core every Applied AI loop opens with.

Retrieval pipelines, chunking, hybrid search, reranking, tool-using agents, guardrails, multi-tenancy and eval harnesses: the modal Applied AI design round at OpenAI, Anthropic, Glean and Sierra.

Practical builds (parsers, in-memory stores, rate limiters, streaming) plus the LeetCode-medium staples, calibrated to the practical coding screens applied AI teams actually run.

Classical ML, embeddings, metrics (precision/recall, AUC, calibration), experiment design and statistics: the modeling fundamentals applied AI roles still test alongside GenAI.

Window functions, gaps-and-islands, Spark internals, lakehouse design, idempotent pipelines and CDC: the data plumbing under every production AI system.

Productionizing a POC: latency and throughput budgets, idempotency and retries, caching, observability for non-deterministic systems, and private or air-gapped deploys.

CI/CD for models, drift detection and retraining, model registries, Kubernetes inference, feature stores, and staging-to-production promotion: what every ML-platform loop drills.

GPU/TPU memory, distributed training and parallelism, quantization, inference serving (vLLM, batching, KV cache), and scaling API gateways: the infra depth the labs and NVIDIA probe.

Prompt-injection defense in depth, PII handling, tenant isolation, abuse and consumption controls, audit trails, and SOC 2 / EU AI Act questions enterprise-facing applied AI work must survive.

Owning ambiguous ML projects end to end, the model-failure post-mortem, cross-functional trade-offs, and the deep-dive on a system you actually shipped: the highest-variance, least-prepped rounds.