[
  {
    "slug": "advanced-kv-caching-disaggregation",
    "name": "Advanced KV Caching & Disaggregation",
    "description": "Prefix caching mechanics, cache-aware routing, KV cache storage tiers, prefill/decode disaggregation, and modern speculative decoding variants.",
    "count": 12
  },
  {
    "slug": "autoscaling-resource-management",
    "name": "Autoscaling & Resource Management",
    "description": "HPA/VPA/Cluster Autoscaler, GPU bin-packing, MIG/fractional GPU sharing, and quota management for AI infrastructure.",
    "count": 10
  },
  {
    "slug": "collective-communication-nccl",
    "name": "Collective Communication & NCCL",
    "description": "Ring vs tree all-reduce, all-gather, reduce-scatter, broadcast, and NCCL fundamentals for GPU cluster communication.",
    "count": 10
  },
  {
    "slug": "container-image-management",
    "name": "Container & Image Management",
    "description": "Docker layer caching, multi-stage builds, and image-size optimization for CUDA/ML container images.",
    "count": 10
  },
  {
    "slug": "cost-optimization-cloud",
    "name": "Cost Optimization & Cloud Infra",
    "description": "Spot vs reserved vs on-demand, GPU utilization economics, and FinOps concepts for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "cuda-programming",
    "name": "CUDA Programming Basics",
    "description": "Kernel launches, the thread/block/grid hierarchy, memory spaces, streams, and the CUDA bugs that show up constantly in practice and in interviews.",
    "count": 10
  },
  {
    "slug": "distributed-training-data-parallel",
    "name": "Distributed Training: Data Parallelism",
    "description": "DDP, gradient synchronization, all-reduce, gradient accumulation, and sync vs async SGD for AI platform engineering interviews.",
    "count": 10
  },
  {
    "slug": "distributed-training-model-parallel",
    "name": "Distributed Training: Model & Pipeline Parallelism",
    "description": "Tensor parallelism, pipeline parallelism, pipeline bubbles, ZeRO stages, FSDP, 3D parallelism, and activation checkpointing.",
    "count": 10
  },
  {
    "slug": "fault-tolerance-checkpointing",
    "name": "Fault Tolerance & Checkpointing",
    "description": "Elastic training, checkpoint/restart strategies, stragglers, and silent data corruption in large-scale training runs.",
    "count": 10
  },
  {
    "slug": "gpu-architecture",
    "name": "GPU Architecture Fundamentals",
    "description": "Core hardware concepts (SMs, warps, tensor cores, and memory hierarchy) that come up in almost every AI platform engineering interview.",
    "count": 10
  },
  {
    "slug": "gpu-memory-bandwidth",
    "name": "GPU Memory & Bandwidth",
    "description": "Coalescing, bank conflicts, the roofline model, and reasoning about whether a kernel is compute-bound or memory-bound.",
    "count": 10
  },
  {
    "slug": "inference-engines-runtime-internals",
    "name": "Inference Engines & Runtime Internals",
    "description": "CUDA kernels, kernel fusion, model file formats, and how vLLM, SGLang, TensorRT-LLM, and NVIDIA Dynamo actually differ under the hood.",
    "count": 13
  },
  {
    "slug": "inference-optimization",
    "name": "Inference Optimization",
    "description": "KV caching, batching strategies, and speculative decoding for fast, high-throughput LLM inference.",
    "count": 10
  },
  {
    "slug": "job-scheduling-orchestration",
    "name": "Job Scheduling & Orchestration",
    "description": "SLURM, Ray clusters, and DAG-based ML pipeline orchestration (Kubeflow Pipelines, Argo Workflows) for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "kubernetes-ml-workloads",
    "name": "Kubernetes for ML Workloads",
    "description": "Scheduling GPU training and serving workloads on Kubernetes: device plugins, taints/tolerations, and the Kubeflow Training Operator.",
    "count": 10
  },
  {
    "slug": "llm-inference-fundamentals",
    "name": "LLM Inference Fundamentals",
    "description": "Model selection, latency metrics, MoE routing, and the roofline model: the foundational judgment calls behind inference engineering.",
    "count": 12
  },
  {
    "slug": "mixed-precision-numerics",
    "name": "Mixed Precision & Numerics",
    "description": "FP32/FP16/BF16/FP8/TF32 tradeoffs, loss scaling, and the numerical-stability questions that come up when discussing large-scale training.",
    "count": 10
  },
  {
    "slug": "mlops-cicd",
    "name": "MLOps & CI/CD",
    "description": "Model registries, experiment tracking, feature stores, deployment strategies, and reproducibility for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "model-serving-frameworks",
    "name": "Model Serving Frameworks",
    "description": "Triton, vLLM, TensorRT-LLM, and TorchServe: how production LLM/model serving stacks are built and scaled.",
    "count": 10
  },
  {
    "slug": "multimodal-inference",
    "name": "Multimodal Inference",
    "description": "Vision-language models, embeddings, ASR, TTS, and image/video generation: inference mechanics and optimization beyond text LLMs.",
    "count": 17
  },
  {
    "slug": "networking-ai-infra",
    "name": "Networking for AI Infrastructure",
    "description": "NVLink, NVSwitch, InfiniBand, RDMA, RoCE, and GPU cluster network topology fundamentals.",
    "count": 10
  },
  {
    "slug": "observability-monitoring",
    "name": "Observability & Monitoring",
    "description": "Metrics, logs, tracing, GPU telemetry (DCGM/Prometheus), and SLOs for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "parallel-computing",
    "name": "Parallel Computing Concepts",
    "description": "The scaling laws, execution models, and shared-memory hazards that underpin every distributed-systems and GPU-performance conversation.",
    "count": 10
  },
  {
    "slug": "production-inference-ops",
    "name": "Production Inference Operations",
    "description": "NIMs, batching strategies, cold starts, scale-to-zero, canary rollouts, GPU procurement, cost estimation, and client-side latency for inference services.",
    "count": 14
  },
  {
    "slug": "profiling-performance-tuning",
    "name": "Profiling & Performance Tuning",
    "description": "Nsight tools, PyTorch profiler, roofline analysis, and diagnosing bottlenecks for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "quantization-compression",
    "name": "Quantization & Model Compression",
    "description": "PTQ vs QAT, GPTQ/AWQ, pruning, and distillation for shrinking models without wrecking accuracy.",
    "count": 10
  },
  {
    "slug": "security-ai-platforms",
    "name": "Security in AI Platforms",
    "description": "Secrets management, supply-chain integrity, multi-tenant isolation, and access control for AI platform interviews.",
    "count": 10
  },
  {
    "slug": "storage-systems-ml",
    "name": "Storage Systems for ML",
    "description": "Object storage vs. parallel filesystems, data-loader I/O bottlenecks, sharding, and checkpoint I/O patterns at training scale.",
    "count": 10
  },
  {
    "slug": "transformer-llm-architecture",
    "name": "Transformer / LLM Architecture Internals",
    "description": "Self-attention math, positional encodings, FFN blocks, and why attention scales quadratically.",
    "count": 10
  },
  {
    "slug": "vector-db-rag-infra",
    "name": "Vector Databases & RAG Infrastructure",
    "description": "ANN search (HNSW/IVF), embedding pipelines, chunking, and retrieval-augmented generation system design.",
    "count": 10
  }
]