First

8. Daily English Learning Plan — AI SRE / Inference Infrastructure #

This companion plan covers 16 weeks × 6 study days = 96 days, at approximately 30 minutes per study day. Day numbers are study sessions, not calendar dates; use the seventh day for rest or catching up. The technical roadmap above remains a separate hands-on commitment. The English weeks follow their own progression and do not have to match the technical roadmap week for week.

Use 3 minutes for recall, 10 minutes for a short English video segment, 8 minutes for approximately 300–500 words of reading, and 9 minutes for the English task. Long talks and documents are reference material: study the indicated topic, not the entire resource. English captions and slower playback are optional supports. Save three reusable phrases each day. Resources do not have official CEFR ratings; adapt the segment length to a B2 learning baseline.

The progression is inference fundamentals → GPU infrastructure → serving → performance and scaling → observability and reliability → quality, cost, and security → architecture decisions. Watching and explaining material does not replace the project completion criteria above.

Contents #

Weeks	Theme	Key Topics	———
1	AI Inference & Production ML Lifecycle	Inference basics, ML lifecycle, MLOps vs SRE, AI stack layers, deployment automation	AI 推理与生产级 ML 生命周期推理基础、机器学习生命周期、MLOps 与 SRE、AI 技术栈分层、部署自动化
2	LLM Fundamentals for Operators	Next-token prediction, transformer flow, attention & context, context window limits, temperature & sampling	运维人员的 LLM 基础下一个 token 预测、Transformer 流程、注意力机制与上下文、上下文窗口限制、温度与采样
3	GPU Hardware, Memory & Runtime	CPU vs GPU, HBM bandwidth, CUDA stack, container GPU access, mixed precision (FP16/FP32)	GPU 硬件、显存与运行时 CPU vs GPU、HBM 带宽、CUDA 栈、容器 GPU 访问、混合精度 (FP16/FP32)
4	GPU Workloads on Kubernetes	Device plugin discovery, GPU Operator, MIG isolation, time-slicing limits, scheduling troubleshooting	K8s 上的 GPU 工作负载设备插件发现、GPU Operator、MIG 隔离、时间片限制、调度排障
5	Model Serving & Inference Engines	Inference runtime role, OpenAI-compatible APIs, Triton, KServe orchestration, TensorRT-LLM	模型服务与推理引擎推理运行时、OpenAI 兼容 API、Triton、KServe 编排、TensorRT-LLM
6	Model Artifacts, Deployment & Startup	Model download & cache, versioning (MLflow), safetensors, OCI packaging, probes for slow loading	模型制品、部署与启动模型下载与缓存、模型版本管理 (MLflow)、safetensors 格式、OCI 镜像打包、慢加载就绪探针
7	Inference Performance & Optimization	Prefill vs decode, KV cache, continuous batching, prefix caching, quantization (GPTQ/AWQ/GGUF)	推理性能与优化 Prefill vs Decode、KV 缓存、连续批处理、前缀缓存、量化 (GPTQ/AWQ/GGUF)
8	Benchmarking Latency, Throughput & Quality	TTFT, ITL, TPS vs RPS, benchmark workload design, quality metrics vs SLOs	延迟、吞吐量与质量基准测试 TTFT、ITL、TPS vs RPS、基准负载设计、质量指标 vs SLO
9	Autoscaling & Capacity Planning	HPA metrics, KEDA scalers, GPU capacity sizing, cost vs latency vs throughput	自动伸缩与容量规划 HPA 指标、KEDA 缩放器、GPU 容量规格选择、成本 vs 延迟 vs 吞吐
10	AI Observability: Metrics, Logs & Traces	vLLM metrics, OpenTelemetry spans, PII redaction, GPU dashboards, prompt drift	AI 可观测性：指标、日志与链路 vLLM 指标、OpenTelemetry Span、PII 脱敏、GPU 仪表盘、Prompt 漂移
11	Inference SLOs, Error Budgets & Alerting	SLO/SLI/SLA, error budgets, AlertManager rules, multi-window burn-rate alerts	推理 SLO、错误预算与告警 SLO/SLI/SLA、错误预算、AlertManager 规则、多窗口消耗率告警
12	Incident Response & Failure Recovery	IR lifecycle, inference triage, blameless postmortem, GPU OOM, chaos engineering	事件响应与故障恢复事件响应生命周期、推理故障分诊、无责事后复盘、GPU OOM、混沌工程
13	Safe Model Releases & Rollback	Model registry stages, canary, blue-green, shadow/dark launch, Argo Rollouts	安全的模型发布与回滚模型仓库阶段划分、金丝雀发布、蓝绿部署、影子/灰度发布、Argo Rollouts
14	RAG Reliability & Quality Evaluation	RAG pipeline, retriever evaluation, RAG triad (relevancy, faithfulness, contextual relevancy)	RAG 可靠性与质量评估 RAG 管线、检索器评估、RAG 三元组（相关性、忠实性、上下文相关性）
15	Cost, Security & Data Protection	Cost per M tokens, GPU right-sizing, spot GPUs, prompt injection, PII redaction	成本、安全与数据保护每百万 token 成本、GPU 合理选型、竞价实例 GPU、提示词注入、PII 脱敏
16	Architecture Trade-offs & Final Review	Throughput vs latency, batch inference, multi-tenancy, build vs buy, hybrid vs self-hosted	架构权衡与总复习吞吐 vs 延迟、批量推理、多租户、自建 vs 外购、混合部署 vs 自托管

Week 1 — AI Inference and the Production ML Lifecycle #

	Day	Topic	Video	Article	English Practice
❇️	01	What happens during AI inference?	AI Inference: The Secret to AI’s Superpowers — IBM Technology — listen for the request-to-prediction flow	What is AI inference? — IBM — inference versus training	Record a 60-second baseline explanation using “input,” “model,” and “prediction.”
	02	How does a model reach production?	Machine Learning Lifecycle Explained — Super Data Science — follow the lifecycle stages	What is the AI lifecycle? — IBM — development and operational feedback	Write 80–100 words using “first,” “next,” and “finally.”
	03	Where do SRE and MLOps overlap?	What is MLOps? — IBM Technology — listen for operational responsibilities	What is MLOps? — IBM — lifecycle automation	Explain the overlap in 90 seconds using “whereas” and “reliability.”
	04	Which layers support an AI service?	Infrastructure Layer: Power the AI Stack — IBM Technology — follow the infrastructure dependencies	What is an AI stack? — IBM — infrastructure, data, and model layers	Write a 100-word dependency map using “depends on” three times.
	05	Why does deployment need automation?	What Is MLOps? — MathWorks, Heather Gorr — focus on production automation	What is model deployment? — IBM — packaging and operational considerations	Explain two deployment risks in 90 seconds using “in order to.”
	06	Review: explain the production ML lifecycle	Machine Learning Engineering for Production — DeepLearning.AI — replay the opening discussion of production challenges	MLOps: Continuous delivery and automation pipelines — Google Cloud — check the lifecycle and ownership	Give a two-minute explanation without notes; compare with Day 01.

Week 2 — LLM Fundamentals for Operators #

	Day	Topic	Video	Article	English Practice
❇️	07	What does an LLM predict?	Large Language Models Explained Briefly — 3Blue1Brown — listen for next-token prediction， Bilibili	Large language models — IBM — definitions and limitations	Explain an LLM in 90 seconds using “predicts” rather than “knows.”
	08	How do tokens pass through a transformer?	Transformers Explained — 3Blue1Brown — follow the opening token workflow， Bilibili	Transformer models — IBM — architecture overview	Write 80–100 words using “first,” “embedding,” and “finally.”
	09	Why does attention use context?	Attention in Transformers — 3Blue1Brown — focus on the first contextual example， Bilibili	Attention mechanisms — IBM — contextual relationships	Explain one example in 90 seconds using “depending on” and “relevant.”
	10	What limits the context window?	What Is the LLM’s Context Window? — New Machina — listen for input and output limits	Context windows — IBM — token limits and operational implications	Write a 100-word explanation using “whereas” and “limit.”
	11	How does sampling change responses?	The Secret Controls for Your LLM — Gary Explains — focus on temperature and sampling	LLM temperature — IBM — randomness versus correctness	Explain in 90 seconds why lower temperature does not guarantee factual accuracy.
	12	Review: explain the operator’s model of an LLM	Intro to Large Language Models — Andrej Karpathy — replay the opening model overview	Large Language Models Explained Briefly — 3Blue1Brown — verify the prediction workflow	Give a two-minute explanation connecting tokens, context, attention, and sampling.

Week 3 — GPU Hardware, Memory, and Runtime Foundations #

	Day	Topic	Video	Article	English Practice
❇️	13	Why use GPUs instead of CPUs for some workloads?	GPUs: Explained — IBM Technology — listen for parallel processing	CPU vs. GPU for machine learning — IBM — architectural trade-offs	Compare CPU and GPU work in 90 seconds using “whereas” twice.
	14	How do memory capacity and bandwidth differ?	High Bandwidth Memory Explained — Techquickie — focus on the memory-interface analogy	High Bandwidth Memory: Everything You Need to Know — Rambus — bandwidth and stacked memory	Write 100 words distinguishing storage capacity from transfer speed.
	15	What does CUDA provide?	Nvidia CUDA in 100 Seconds — Fireship — listen for the software and hardware relationship	CUDA platform — NVIDIA — compiler, libraries, and runtime	Explain the CUDA stack in 90 seconds using “enables” and “depends on.”
	16	How does a container access a GPU?	Docker with NVIDIA GPU Support — Jeremy Pedersen — follow the runtime integration concept	NVIDIA Container Toolkit overview — runtime components	Write 100 words separating the host driver from container libraries; do not copy old installation commands.
	17	Why does numerical precision matter?	Mixed Precision Training — Aladdin Persson — focus on the introductory memory and speed explanation	Train with mixed precision — NVIDIA — half precision and numerical trade-offs	Explain FP16 versus FP32 in 90 seconds; distinguish the training example from inference.
	18	Review: describe the GPU execution stack	GPU Architecture Deep Dive — Parallel Routines — replay the opening hardware overview	GPU Performance Background User’s Guide — NVIDIA — verify compute and memory terminology	Give a two-minute explanation connecting cores, memory, runtime, and workload.

Week 4 — Operating GPU Workloads on Kubernetes #

	Day	Topic	Video	Article	English Practice
❇️	19	How does Kubernetes discover GPUs?	GPUs on Kubernetes: What Actually Happens — CNCF, Topcu and Polencic — follow the device allocation path	Device plugins — Kubernetes — discovery and resource reporting	Explain device discovery in 90 seconds using “advertises,” “allocates,” and “kubelet.”
	20	What does the GPU Operator manage?	Mastering GPU Management Using the Operator Pattern — CNCF — focus on the managed components	About the NVIDIA GPU Operator — driver, toolkit, and plugin responsibilities	Write 100 words explaining which responsibilities the operator automates.
	21	What isolation does MIG provide?	Efficient Access to Shared GPU Resources — CNCF — listen for physical partitioning	GPU Operator with MIG — NVIDIA — supported hardware and partitions	Compare a full GPU and a MIG instance in 90 seconds using “whereas.”
	22	What are the limits of time-slicing?	GPU Sharing and CDI in Device Plugins — CNCF — focus on sharing mechanisms	Time-slicing GPUs in Kubernetes — NVIDIA — memory and fault-isolation limitations	Write 100–120 words explaining why a shared slot is not a dedicated memory partition.
	23	Why might a GPU Pod remain pending?	Scheduled GPUs on Kubernetes — Eric O Meehan — follow the scheduling setup	Schedule GPUs — Kubernetes — requests, limits, and node selection	Describe three scheduling checks in 90 seconds using “first,” “then,” and “if.”
	24	Review: explain a GPU workload from request to execution	Kubernetes with GPUs Like I’m 5 — CNCF, Carlos Santana — replay the orchestration overview	Installing the GPU Operator — NVIDIA — check prerequisites and validation steps	Give a two-minute walkthrough covering discovery, scheduling, runtime, and sharing.

Week 5 — Model Serving and Inference Engines #

	Day	Topic	Video	Article	English Practice
❇️	25	What does an inference runtime do?	What an Inference Runtime Actually Does — Mahesh Dsouza — follow the serving responsibilities	Quickstart — vLLM — offline inference and online serving	Explain the runtime’s role in 90 seconds using “loads,” “schedules,” and “generates.”
	26	What does an OpenAI-compatible endpoint provide?	Understanding vLLM with a Hands-On Demo — KodeKloud — focus on the API server example	OpenAI-compatible server — vLLM — supported APIs and limitations	Write 100 words explaining why compatible APIs do not imply identical behavior.
	27	What does Triton Inference Server manage?	Top 5 Reasons Why Triton Is Simplifying Inference — NVIDIA Developer — listen for model-serving capabilities	Triton Inference Server overview — NVIDIA — serving across model frameworks	Explain Triton’s role in 90 seconds using “backend” and “endpoint.”
	28	How does KServe fit above a model server?	Deploy Models with KServe, MLServer and MLflow — MLWorks — focus on component responsibilities	Deploy your first predictive InferenceService — KServe — declarative serving	Write 100 words distinguishing orchestration from model execution.
	29	What does TensorRT LLM optimize?	TensorRT LLM Introduction — Fahd Mirza — listen for acceleration goals	TensorRT LLM overview — NVIDIA — optimization scope and capabilities	Explain a possible benefit and compatibility constraint in 90 seconds.
	30	Review: choose the layers of a serving stack	AI Infrastructure Explained — KodeKloud — replay the serving-stack overview	What is AI infrastructure? — IBM — check infrastructure responsibilities	Give a two-minute recommendation; distinguish a runtime, a server, and orchestration.

Week 6 — Model Artifacts, Deployment, and Startup #

	Day	Topic	Video	Article	English Practice
❇️	31	How are model files downloaded and cached?	Download a Model from Hugging Face — Amit Thinks — follow the artifact retrieval workflow	Download files from the Hub — Hugging Face — revisions and caching	Explain a download failure and a cache hit in 90 seconds.
	32	Why record the model version?	MLflow Model Tracking and Model Registry — FourthBrainAI — listen for lineage and registration	ML model registry — MLflow — versions, aliases, and metadata	Write a 100-word release record using “version,” “artifact,” and “reproduce.”
	33	Which files make a model usable?	Hugging Face AI Model Files Explained — Fahd Mirza — distinguish weights, configuration, and tokenizer files	Safetensors security audit — Hugging Face — safer tensor serialization and its scope	Explain in 90 seconds why safe weights do not make all accompanying code trustworthy.
	34	What belongs in a deployable model package?	Machine Learning Packaging for Ops with KitOps — Bret Fisher — focus on reproducible bundles	ModelKit overview — KitOps — OCI packaging for model artifacts	Write 100 words comparing a model package with a container image.
	35	When is a model server ready for traffic?	Liveness vs. Readiness vs. Startup Probes — Anton Putra — listen for the three probe purposes	Configure liveness, readiness, and startup probes — Kubernetes — startup protection and readiness	Explain how slow model loading affects probe design in 90 seconds.
	36	Review: trace a model artifact into a ready endpoint	KitOps and KServe Integration Guide — Jozu — replay the deployment workflow	Deploy your first LLM with InferenceService — KServe — check the model-to-endpoint path	Give a two-minute walkthrough covering revision, download, loading, and readiness.

Week 7 — Inference Performance and Optimization #

	Day	Topic	Video	Article	English Practice
❇️	37	What is prefill versus decode?	Why LLMs Read Fast but Write Slowly - Prefill vs Decode — ML Guy — listen for the two-phase timing difference	Prefill and Decode for Concurrent Requests — Hugging Face — latency budgets under load	Explain in 90 seconds why time-to-first-token and inter-token latency behave differently.
	38	What does the KV cache store?	KV Cache Explained: Speed Up LLM Inference with Prefill and Decode — Ready Tensor — follow the key-and-value reuse example	Paged Attention — vLLM — how attention memory is partitioned	Write 100 words explaining why a KV cache saves compute but uses GPU memory.
	39	How does continuous batching raise throughput?	Continuous Batching for LLM Inference — Uplatz — focus on iteration-level scheduling	Continuous batching from first principles — Hugging Face — static versus dynamic batching	Explain in 90 seconds why continuous batching reduces the padding waste of static batches.
	40	What does prefix caching reuse?	What is Prefix Caching? — Standarity — listen for the shared-prefix reuse story	Automatic Prefix Caching — vLLM — enabling APC and limits	Write 100 words describing one workload where prefix caching helps and one where it does not.
	41	Why quantize LLM weights?	LLM Quantization Techniques Explained — Joydeep Bhattacharjee — focus on the GPTQ, AWQ, and GGUF comparison	Quantization — vLLM — supported formats and quality trade-offs	Explain in 90 seconds why INT4 quantization trades numerical precision for memory and speed.
	42	Review: weigh inference-performance trade-offs	Mastering LLM Inference Optimization From Theory to Cost Effective Deployment — AI Engineer, Mark Moyou — replay the opening cost framework	LLM Inference at scale with TGI — Hugging Face — check the end-to-end optimization stack	Give a two-minute walkthrough linking prefill, KV cache, batching, prefix caching, and quantization to one cost scenario.

Week 8 — Benchmarking Latency, Throughput, and Quality #

	Day	Topic	Video	Article	English Practice
❇️	43	What is time to first token?	LLM Inference Performance: Latency and Throughput Metrics — Ready Tensor — listen for the TTFT definition	Metrics — NVIDIA NIM LLMs Benchmarking — TTFT, e2e latency, ITL, TPS, RPS	Explain in 90 seconds which stages of inference contribute to TTFT.
	44	What is inter-token latency?	RTX 4090 vLLM Benchmark — Database Mart — focus on the latency numbers	Benchmark CLI — vLLM — benchmark serving and throughput scripts	Write 100 words describing how to record ITL under varying concurrency.
	45	How do TPS and RPS compare?	GuideLLM: Evaluate your LLM Deployments for Real-World Inference — Red Hat, vLLM Office Hours — listen for the per-system versus per-user framing	Understand LLM latency and throughput metrics — Anyscale — tokens, requests, and concurrency	Explain in 90 seconds why total TPS can grow while TPS per user shrinks.
	46	How do you design a realistic benchmark workload?	Ultimate Guide to LLM Benchmarks — Bhavesh Bhatt — focus on the workload-shape examples	LLM inference latency — ClickHouse — capturing TTFT and TPOT in practice	Write 100 words listing the workload variables a benchmark report must record.
	47	How is model quality measured?	Hallucination Rate Explained in AI Testing — CodeCraft Academy — listen for the quality-metric framing	LLM Benchmarks Explained — Atlan — what benchmark scores actually mean	Explain in 90 seconds why one aggregate score cannot replace service SLOs.
	48	Review: tie latency, throughput, and quality together	Tour De Force: LLM Inference Optimization From Simple To Sophisticated — PyTorch, Christin Pohl, Microsoft — replay the optimization layering	Benchmark Suites — vLLM — check benchmark types and outputs	Give a two-minute walkthrough picking a metric set for one documented workload.

Week 9 — Autoscaling and Capacity Planning #

	Day	Topic	Video	Article	English Practice
❇️	49	What does HPA scale on?	Kubernetes HPA Explained in 10 Minutes — Cloud Guru — listen for the metric, target, and stabilization window	Horizontal Pod Autoscaling — Kubernetes — metrics, behaviors, and policies	Explain in 90 seconds how a target utilization becomes a replica count.
	50	Why is CPU-based HPA unreliable for LLMs?	Optimizing Load Balancing and Autoscaling for LLM Inference on Kubernetes — CNCF, D. Gray — focus on the queue and prefill pressures	Best practices for autoscaling LLM inference — Google Cloud — metric selection for inference	Write 100 words explaining why queue depth or pending requests beat CPU for LLM scale-out.
	51	How does KEDA extend HPA?	Run Scalable LLMs on Kubernetes in Minutes (Using KEDA) — DigitalOcean — follow the ScaledObject definition	Autoscaling with KEDA — vLLM Production Stack — Prometheus trigger and queue threshold	Explain in 90 seconds how KEDA scales a Deployment from a Prometheus query.
	52	What triggers a KEDA scaler?	KEDA: Kubernetes Event-Driven Autoscaling — DevOps & AI Toolkit — listen for the scaler and polling examples	Autoscale GPU and LLM workloads on Kubernetes — Kedify — inference-aware routing and HTTP scalers	Write 100 words comparing KEDA scalers with custom-metric HPA.
	53	How do you size GPU capacity?	GPU Instance Selection: AI & LLM Inference Benchmarking — Sam Mokhtari — focus on the memory-versus-cost framework	Inference Is the New Bottleneck: Plan GPU Capacity for Production AI — ClearML — workload-driven capacity sizing	Explain in 90 seconds why model size, KV cache, and concurrency drive GPU count.
	54	Review: connect autoscaling to capacity	LLM’s INFERENCE: Cost vs. Latency vs. Throughput — César Soto Valero — replay the cost-versus-latency framing	Cost-Efficient AI Inference Cloud Strategies — GMI Cloud — workload shapes and reserved capacity	Give a two-minute walkthrough sizing replicas, headroom, and scale-down safety.

Week 10 — AI Observability: Metrics, Logs, and Traces #

	Day	Topic	Video	Article	English Practice
❇️	55	What does vLLM expose as metrics?	Get More Performance From Your DGX Spark — vLLM + Grafana Tuning Dashboard — Ryan Susman — follow the metric-to-panel example	Metrics — vLLM — vLLM engine and request metrics	Explain in 90 seconds which vLLM metric best reflects user-perceived queueing.
	56	How does OpenTelemetry trace an LLM call?	Intro to OpenTelemetry and LLM Observability — Arize AI — listen for the span and attributes example	OpenTelemetry Concepts — signals, instrumentation, context	Write 100 words listing the spans you would record across one LLM request.
	57	How do you log prompts safely?	WSO2 AI Guardrails: PII Masking, Prompt Injection & Safety — WSO2 — focus on the masking and redaction steps	Redact PII from LLM Telemetry Without Losing Debuggability — OpenObserve — masking trade-offs	Explain in 90 seconds why redaction is required before sending traces to a vendor.
	58	What belongs on a GPU dashboard?	How to Monitor Key LLM Metrics (GPU + Grafana Dashboard) — Saujan Bohara — follow the panel selection	Enabling the GPU Monitoring Dashboard — NVIDIA GPU Operator — DCGM exporter metrics	Write 100 words describing the panels you would keep, drop, or add for one inference cluster.
	59	How is prompt drift detected?	Drift Monitoring and Evaluation for LLM Apps — Evidently AI, MG — listen for the reference-baseline framing	Introducing Evidently 0.0.1: Open-Source Tool To Analyze Data Drift — Evidently AI — reference versus current data	Explain in 90 seconds why drift is a signal to investigate, not a retraining trigger.
	60	Review: tie metrics, logs, and traces together	Observability and Its Pillars Explained — OpenObserve — replay the three-pillar framing	Three Pillars of Observability — IBM — logs, metrics, and traces	Give a two-minute walkthrough showing which pillar answers which operational question.

Week 11 — Inference SLOs, Error Budgets, and Alerting #

	Day	Topic	Video	Article	English Practice
❇️	61	What is an SLO?	SLO vs SLI vs SLA vs Error Budget — Tech Tutorials with Piyush — listen for the reliability-target framing	Designing SLOs — Google Cloud — SLI types, compliance, and error budget	Explain in 90 seconds the difference between SLO, SLI, and SLA.
	62	Which metric makes a good SLI?	SLO & SLI Explained: Service Level Objectives & Indicators for Beginners — CodeLucky — follow the SLI-selection examples	SLA vs SLO vs SLI Explained: Key Differences — NovelVista — what each term really covers	Write 100 words proposing SLIs for one documented inference workload.
	63	What is an error budget?	Error Budgets Explained: Balance Innovation & Reliability — CodeLucky — listen for the budget-versus-risk framing	How to Set SLOs in 2026: SLIs, Error Budgets & Burn-Rate Alerts — OpenObserve — common SLO mistakes	Explain in 90 seconds how an error budget connects availability and release velocity.
	64	How do you write an alert rule?	Full Tutorial: AlertManager Set up and PrometheusRules — Anais Urlichs — follow the rule and routing example	Configuration — Prometheus — alerting rules, routes, and receivers	Write 100 words distinguishing a symptom alert from a budget-burn alert.
	65	What is multi-window burn-rate alerting?	Defending SLOs: Error Budget Burn Rate Alerting using Datadog — DevOps Austin — listen for the window-pair logic	How to Build Multi-Burn-Rate SLO Alerts from OpenTelemetry Metrics — OneUptime — burn-rate recording and alert rules	Explain in 90 seconds why a single threshold is not enough for an SLO alert.
	66	Review: connect SLOs, budgets, and alerts	SREcon15 — Error Budgets and Risks — USENIX — replay the budget-versus-risk framing	How to implement multi-window, multi-burn-rate alerts with Grafana Cloud — window-pair construction	Give a two-minute walkthrough defining one SLI, one SLO, and one alert tier.

Week 12 — Incident Response and Failure Recovery #

	Day	Topic	Video	Article	English Practice
❇️	67	What is an incident response lifecycle?	The 6 Steps of the Incident Response Life Cycle — Cyber Gray Matter — follow the detect-to-learn sequence	Incident Response Process: Step-by-Step Guide for SRE Teams — Rootly — lifecycle stages and ownership	Explain in 90 seconds which stages an SRE typically owns and which need a separate responder.
	68	How do you triage an inference failure?	How Site Reliability Engineering Handles High-Severity Incidents — SystemDR — listen for the triage and command example	Troubleshooting GPU Memory Out-of-Memory Errors — NVIDIA NIM — OOM symptoms and mitigations	Write 100 words listing the first three signals you would check for a slow LLM endpoint.
	69	What is a blameless postmortem?	Blameless Post Mortems — Code for America, John Allspaw — listen for the systems-versus-people framing	How to run a blameless postmortem — Atlassian — story, contributing factors, and actions	Explain in 90 seconds why a postmortem lists contributing factors instead of human faults.
	70	What causes GPU OOM during inference?	Learn vLLM: Troubleshooting Deepseek R1 8B GPU OOM on single L4 GPU — Samos123 — follow the OOM investigation	vLLM OOM Errors: Root Cause Diagnosis Guide — Paralleliq — KV cache, batch size, fragmentation	Write 100 words distinguishing a model-size OOM from a KV-cache OOM.
	71	How does chaos engineering help inference reliability?	Chaos Engineering Explained: How to Build Failure-Resilient Systems — Harness — focus on hypothesis and steady state	What is Chaos Engineering? — IBM — controlled failure injection	Explain in 90 seconds why chaos tests run in a lab first and what evidence they should produce.
	72	Review: connect detection, response, and learning	What is an AI SRE? — ResolveAI — replay the detect-to-learn automation framing	The AI Agent Incident Response Runbook — Reco — containment, analysis, recovery	Give a two-minute walkthrough of one incident from alert to postmortem action item.

Week 13 — Safe Model Releases and Rollback #

	Day	Topic	Video	Article	English Practice
❇️	73	How do you version and stage a model?	MLflow Model Registry Explained — YourMLStudents — listen for the staging-to-production transition	Manage model lifecycle using the Workspace Model Registry — Databricks — lifecycle stages and ownership	Explain in 90 seconds why staging and production are separate steps.
	74	What is a canary release?	Canary Deployments Explained — Harness — focus on the gradual traffic shift	Canary Deployment Strategy — Argo Rollouts — steps, weights, and pause conditions	Write 100 words explaining why canary shifts allow an early rollback.
	75	What is blue-green deployment?	Kubernetes Blue Green Deployment Strategy Explained — Aman Pathak — follow the active and preview service split	BlueGreen Deployment Strategy — Argo Rollouts — active, preview, and scale-down delay	Explain in 90 seconds why blue-green gives instant rollback but doubles resource use briefly.
	76	What is a dark launch or shadow deployment?	How Do Dark Launches Work In Continuous Deployment? — Cloud Stack Studio — listen for the mirrored-traffic pattern	Kubernetes Deployment Strategies — Octopus Deploy — rolling, blue-green, canary, and shadow	Write 100 words describing what shadow deployment tells you that canary cannot.
	77	How do you automate rollouts in Kubernetes?	Argo Rollouts — Canary Deployments Made Easy In Kubernetes — DevOps & AI Toolkit — follow the Rollout CRD example	Argo Rollouts — Kubernetes Progressive Delivery — controller, CRDs, and traffic management	Explain in 90 seconds how a Rollout differs from a Deployment in steady state.
	78	Review: pick a safe release for one scenario	Progressive Delivery Explained — DevOps & AI Toolkit — replay the rollout-strategy comparison	Analysis — Argo Rollouts — background analysis that gates a promotion	Give a two-minute walkthrough picking one release strategy and one guardrail for a documented inference workload.

Week 14 — RAG Reliability and Quality Evaluation #

	Day	Topic	Video	Article	English Practice
❇️	79	What is RAG?	What is Retrieval-Augmented Generation (RAG)? — IBM Technology — follow the retrieve-then-generate flow	What is RAG? — IBM — knowledge base, retriever, generator	Explain in 90 seconds why retrieval can improve factuality without retraining.
	80	What are the parts of a RAG pipeline?	RAG Explained in 10 Minutes — Cloud Quick Labs — listen for chunking and embedding	What is Retrieval-Augmented Generation (RAG)? — Databricks — ingestion, retrieval, augmentation, generation	Write 100 words naming the offline ingestion path and the online query path.
	81	How do you build a RAG pipeline?	RAG Tutorial 2026 #1 — Harish Neel — focus on the end-to-end walkthrough	Faithfulness — Ragas — factual consistency of the response with retrieved context	Explain in 90 seconds why a faithful answer still may not be relevant.
	82	How do you evaluate the retriever?	Vector Databases Explained — Aishwarya Srinivasan — focus on similarity search and chunking trade-offs	Understanding RAG Part IV: RAGAs & Other Evaluation Frameworks — Machine Learning Mastery — contextual precision and recall	Write 100 words distinguishing recall from precision in retrieval.
	83	How do you evaluate the generated answer?	LLM Evaluation Explained — Schovia — follow the accuracy, faithfulness, and hallucination framing	Using the RAG Triad for RAG evaluation — DeepEval — answer relevancy, faithfulness, contextual relevancy	Explain in 90 seconds why the RAG triad is referenceless.
	84	Review: connect retrieval, generation, and metrics	RAG MasterClass — Build A RAG Pipeline From Scratch — A.I Engineering BootCamp — replay the architecture recap	RAG Evaluation Metrics — Confident AI — answer relevancy, faithfulness, and more	Give a two-minute walkthrough tracing one user query through retrieval, generation, and the three RAG triad metrics.

Week 15 — Cost, Security, and Data Protection #

	Day	Topic	Video	Article	English Practice
❇️	85	What is the cost per million tokens?	LLM Cost Optimization: Token Economics for Agentic FinOps — KryptoMindz Technologies — follow the cost-per-token framing	AI Inference Cost Economics in 2026 — Spheron — cost per million tokens and four optimization layers	Explain in 90 seconds why CPM is the most useful cost metric for inference.
	86	How do you right-size GPU memory?	How Much GPU Memory is Needed for LLM Inference? — AppliedAI — listen for the model-size plus KV-cache formula	Optimizing Inference Costs: The Complete Guide — Mirantis — right-sizing and idle-capacity avoidance	Write 100 words explaining why model weights, KV cache, and concurrency all fit into the sizing math.
	87	Are spot GPUs worth the risk?	Running Multiple Models on the Same GPU, on Spot Instances — Toronto Machine Learning Society — focus on interruption handling	What Are Spot GPUs? — Northflank — discounts, interruption warnings, and fit for inference	Explain in 90 seconds why spot GPUs change the answer to “where does state live?”
	88	What is prompt injection?	What Is a Prompt Injection Attack? — IBM Technology — listen for direct and indirect injection	LLM01:2025 Prompt Injection — OWASP Gen AI Security Project — risk description and mitigations	Write 100 words distinguishing direct injection from indirect injection through retrieved content.
	89	How do you redact PII before telemetry?	Protect Sensitive Data with Azure AI Language PII Redaction — Microsoft Developer — follow the redaction pipeline	LLM Security Playbook — Kong — prompt injection, data leaks, and model theft	Explain in 90 seconds why redaction has to happen before traces leave your perimeter.
	90	Review: tie cost, security, and privacy together	AI and Data Privacy Explained — Privacy Trainer — replay the GDPR and AI Act framing	How Much Does Your LLM Inference Cost? — NVIDIA Technical Blog — TCO formula and benchmarking inputs	Give a two-minute walkthrough showing how one cost-saving choice affects security and privacy controls.

Week 16 — Architecture Trade-offs and Final Review #

	Day	Topic	Video	Article	English Practice
❇️	91	How do throughput and latency trade off?	ML Model Serving – Latency vs Throughput — Uplatz — follow the latency-versus-throughput framing	The LLM Inference Trilemma — DigitalOcean — throughput, latency, and cost trade-offs	Explain in 90 seconds why per-user latency can rise while total throughput grows.
	92	When is batch inference the right choice?	Scaling LLM Batch Inference: Ray Data & vLLM — InfoQ — focus on the batch pipeline	Realtime vs Batch Inference — Inworld AI — serving contracts, latency budgets, and batching wins	Write 100 words naming one workload that fits batch inference and one that does not.
	93	How do you serve multiple tenants safely?	Multi-Tenant AI-as-a-Service for AI Factories — OpenNebula — listen for isolation and quota	Multi-Tenant LLM Serving on GPU Cloud — Spheron — per-customer quotas and noisy-neighbor isolation	Explain in 90 seconds why request quotas are not the same as token quotas.
	94	Build versus buy: when is an API the right answer?	API vs Self-Hosted LLMs — CodeLucky — follow the control-versus-convenience comparison	Managed LLM API vs Self-Hosting — Access All GPT — decision criteria, cost, and compliance	Write 100 words listing two scenarios where a managed API wins and two where self-hosting wins.
	95	Hybrid or fully self-hosted?	Self-Hosting LLMs: Architect’s Guide to When & How — InfoQ — focus on the decision framework	Hybrid Cloud vs On-Premise LLM Deployment — newline — routing, gateway, and cost trade-offs	Explain in 90 seconds why hybrid setups add a routing layer that pure cloud or on-prem does not.
	96	Final review: connect every week’s lesson	Supercharging IT Operations SRE Agents, GPU Metrics Optimization, and Capacity Risk Management — BMC Helix Events — replay the AI-SRE framing	What is an AI SRE? The Complete AI SRE Guide for 2026 — Rootly — incident response, root cause, and remediation	Give a two-minute walkthrough connecting lifecycle, hardware, serving, performance, observability, reliability, RAG, cost, security, and architecture into one AI-SRE checklist.

Weekly Self-Check #

After each sixth study day, record:

Listening: Can I identify the main point and three supporting details without captions?
Vocabulary: Can I use five useful phrases from this week in new sentences?
Speaking: Can I explain the topic for two minutes using only a few keywords?
Writing: Can I produce a short summary with a main point, an example, and a limitation?
Next step: Which one difficulty should I focus on next week? Repeat a difficult week if needed; finishing on schedule is not a language-proficiency test.