Module 01: Master Map
Course Overview & 45 Scenario Matrix
🎯 Curriculum Overview & Master Matrix
This 11-module masterclass translates 50 real-world technical debriefs, failure war-stories, and architectural whiteboard challenges collected from senior AI candidates and leads at OpenAI, Google, NVIDIA, AWS, and Netflix into an interactive learning syllabus.
🔍 Interactive Scenario Matrix Filter
Filter all 45 interview scenarios by domain or role track:
| ID | Domain | Role Track | Core Scenario / Interview Question | Difficulty | Module Page |
|---|---|---|---|---|---|
| QA-01 | Production Failure | AI DevOps | The "Tell me about a 2 AM failure" in production AI | Hard | 09. SRE & Incidents |
| QA-02 | System Design | AI Engineer | Whiteboard: End-to-End Production RAG System | Hard | 02. RAG & Search |
| QA-03 | Agentic Workflows | AI Engineer | LangGraph Checkpointer Crash & Duplicate Payment Bug | Hard | 05. Agentic AI |
| QA-04 | GPU Architecture | AI DevOps | GPU Utilization vs Occupancy vs Memory Bandwidth | Hard | 04. Memory Math |
| QA-05 | Inference Serving | AI DevOps | Continuous Batching & PagedAttention in vLLM | Hard | 03. Serving & Kernels |
| QA-06 | Memory Sizing | AI DevOps | Llama-3-70B VRAM & GQA KV-Cache Math (64 streams) | Hard | 04. Memory Math |
| QA-07 | Autoscaling | AI DevOps | Why HPA CPU/GPU utilization fails for LLM serving | Medium-Hard | 08. K8s & Autoscaling |
| QA-08 | Cloud Resiliency | Architect | "What happens when an entire cloud region fails?" | Hard | 09. SRE & Incidents |
| QA-09 | Troubleshooting | AI DevOps | Intermittent 502 Bad Gateways under AI Inference load | Medium-Hard | 09. SRE & Incidents |
| QA-10 | High-Traffic Incident | AI DevOps | "Every morning at 10 AM, users see 'Something went wrong'" | Medium-Hard | 09. SRE & Incidents |
| QA-11 | Cost & Routing | AI Engineer | Intelligent Model Routing (SLMs -> LLMs) | Medium-Hard | 03. Serving & Kernels |
| QA-12 | Memory Exhaustion | AI DevOps | Cluster at 92% HBM; KV cache at 60%; cannot buy GPUs | Hard | 04. Memory Math |
| QA-13 | K8s Networking | AI DevOps | 5,000 Services & kube-proxy at 100% CPU in AI cluster | Hard | 08. K8s & Autoscaling |
| QA-14 | AWS / CI/CD | AI DevOps | GitHub Actions OIDC to AWS failing with AccessDenied | Medium-Hard | 09. SRE & Incidents |
| QA-15 | RAG Failure Triage | AI Engineer | Model returns bad answer: Dissecting 5 root causes | Hard | 02. RAG & Search |
| QA-16 | Agent Reliability | AI Engineer | Agent Infinite Tool Loops & Idempotent Retries | Medium-Hard | 05. Agentic AI |
| QA-17 | Context Tradeoffs | AI Engineer | "If models have 1M+ tokens, why not pass all docs?" | Medium | 02. RAG & Search |
| QA-18 | Security & Escapes | AI DevOps | Indirect Prompt Injection & Code Sandboxing | Hard | 10. Evals & Drills |
| QA-19 | Evaluation CI/CD | MLOps | Automated Quality Gates (Ragas, Faithfulness, Judges) | Medium-Hard | 10. Evals & Drills |
| QA-20 | HPC & Interconnect | AI DevOps | InfiniBand vs RoCE & NCCL Topology for Multi-Node | Hard | 07. Distributed Scale |
| QA-21 | Cold Starts | AI DevOps | Reducing 70B model cold start from 8m to <45s | Medium-Hard | 08. K8s & Autoscaling |
| QA-22 | IaC Drift | AI DevOps | Terraform plan wants to destroy un-touched state | Medium | 09. SRE & Incidents |
| QA-23 | Vector DB Scaling | MLOps | 50M Vectors: HNSW vs IVF-PQ & Preventing OOM Kills | Hard | 02. RAG & Search |
| QA-24 | Tokenomics | AI Engineer | Byte-Pair Encoding failure on numbers and code | Medium | 06. Fine-Tuning |
| QA-25 | Distributed Parallelism | AI DevOps | Tensor (TP) vs Pipeline (PP) vs Data Parallelism (DP) | Hard | 07. Distributed Scale |
| QA-26 | Agent Protocols | AI Engineer | Model Context Protocol (MCP) vs Traditional REST APIs | Medium-Hard | 05. Agentic AI |
| QA-27 | Kernel Optimization | AI Engineer | FlashAttention Tiling & Online Softmax Algorithm | Hard | 03. Serving & Kernels |
| QA-28 | Fine-Tuning vs PEFT | AI Engineer | LoRA vs QLoRA Mathematical Formulation & S-LoRA | Hard | 06. Fine-Tuning |
| QA-29 | Format Failure | AI Engineer | Prompting vs Fine-Tuning Decision Tree | Medium-Hard | 06. Fine-Tuning |
| QA-30 | Distributed Training | AI DevOps | DeepSpeed ZeRO-1/2/3 Memory Partitioning Math | Hard | 07. Distributed Scale |
| QA-31 | Serving Asymmetry | AI Engineer | Speculative Decoding & The Core Transformer Asymmetry | Hard | 03. Serving & Kernels |
| QA-32 | Scale Architecture | Architect | Prefill-Decode Disaggregation (PD Disaggregation) | Hard | 03. Serving & Kernels |
| QA-33 | Distributed Serving | AI DevOps | Ray Serve + vLLM High-Throughput Cluster Architecture | Medium-Hard | 07. Distributed Scale |
| QA-34 | Quantization Showdown | MLOps | AWQ vs GPTQ vs SmoothQuant vs GGUF Tradeoffs | Hard | 06. Fine-Tuning |
| QA-35 | Runtime Security | AI DevOps | eBPF Runtime Threat Auditing & Zero-Downtime CVE Patch | Hard | 10. Evals & Drills |
| QA-36 | Serving Architecture | AI DevOps | Triton Inference Server Multi-Model Pipelines & SLAs | Hard | 03. Serving & Kernels |
| QA-37 | Advanced Search | Architect | GraphRAG vs Vector RAG: Multi-Hop Macro Queries | Hard | 02. RAG & Search |
| QA-38 | Agent Security | AI DevOps | Indirect Prompt Injection & Dual-LLM Privilege Isolation | Hard | 10. Evals & Drills |
| QA-39 | Interconnect | AI DevOps | InfiniBand vs RoCE v2 & NCCL Step-Time Jitter Triage | Hard | 07. Distributed Scale |
| QA-40 | Distributed K8s | MLOps | Gang Scheduling with Volcano & Preventing Resource Deadlocks | Hard | 08. K8s & Autoscaling |
| QA-41 | Quality & CI/CD | MLOps | Automated LLM Eval Quality Gates (Ragas/Promptfoo) | Medium-Hard | 10. Evals & Drills |
| QA-42 | Fine-Tuning Data | AI Engineer | Synthetic Data Distillation & Multi-Tier Execution Filter | Medium-Hard | 06. Fine-Tuning |
| QA-43 | Cost Arbitrage | Architect | Intelligent Model Routing & Semantic Caching Gateway | Medium-Hard | 03. Serving & Kernels |
| QA-44 | Performance Profiling | AI DevOps | vLLM Chunked Prefill, PagedAttention & ITL Stutter Triage | Hard | 03. Serving & Kernels |
| QA-45 | Fault Tolerance | Architect | Zero-Downtime Migration of Long-Running Agent Workflows | Hard | 05. Agentic AI |
| QA-46 | Frontier Modeling | AI Engineer | Test-Time Compute Scaling, o1/R1, PRM vs ORM Tree Search | Hard | 11. Reasoning & Triton |
| QA-47 | Post-Training Alignment | AI Engineer | DPO vs PPO vs GRPO: Eliminating the Critic Network | Hard | 11. Reasoning & Triton |
| QA-48 | Kernel Optimization | AI DevOps | Writing Custom Fused Kernels in OpenAI Triton (SRAM Tiling) | Hard | 11. Reasoning & Triton |
| QA-49 | Quantization Formats | AI DevOps | FP8 (E4M3 vs E5M2) vs INT4 vs AWQ Tensor Core GEMM | Hard | 11. Reasoning & Triton |
| QA-50 | Frontier Inference | Architect | Speculative Decoding & Multi-Token Prediction (MTP) Engine | Hard | 11. Reasoning & Triton |