8-Week MLOps + AIOps Bootcamp on AWS (Real-World Projects)

8 Weeks, 16 Classes, 180 min each
Saturday & Sunday, 12 PM – 3 PM IST
Instructor: Akhilesh Mishra | Live Classes (English) | Lifetime Access

By the end of this bootcamp, you will have built a Production RAG & AI Platform on AWS. It covers ML model pipelines with drift detection, serverless RAG on Bedrock, AI agents built with Strands, full AI observability with OpenTelemetry and Langfuse, and self-service AI infrastructure delivered through Terraform and GitHub Actions.

Prerequisites: AWS, Docker, Python, Kubernetes, and Terraform basics.

Module 1: MLOps (3 Weeks)

Week 1: ML Foundations for Platform Engineers

Class 1: How ML Reaches Production

• The ML lifecycle: data → training → evaluation → serving → monitoring
• What platform engineers own and what data scientists own
• Training a simple model locally to understand what you’ll operate
• MLflow experiment tracking: params, metrics, artifacts

Class 2: MLflow on Kubernetes + Data Versioning

• MLflow on EKS with an S3 artifact store and an RDS backend, provisioned with Terraform
• MLflow model registry: versions, aliases, promotion workflow
• DVC for data versioning, treating datasets like code

Project: Run MLflow on EKS via Terraform, track 3 training runs, version the dataset with DVC, and promote the best model in the registry.

Week 2: Training Pipelines + Model Serving

Class 3: Training Platform on EKS

• Kubeflow Pipelines vs Argo Workflows
• DAG workflows: parameterized runs and cached steps
• GPU nodes on EKS, with Spot via Karpenter and on-demand fallback

Class 4: Serving Models in Production

• Serving patterns (batch, real-time, async) and when to use each
• BentoML on EKS vs SageMaker endpoints: the honest trade-off
• KEDA autoscaling on SQS queue depth

Project: Build a training pipeline (validate → train → evaluate → register) on Spot, deploy the model on EKS and on SageMaker, and compare cost and latency.

Week 3: ML CI/CD + Monitoring + Drift Detection

Class 5: ML CI/CD with GitHub Actions

• Champion vs challenger: promote only if the new model wins
• Canary model rollout with Istio at 10% traffic
• An automated ML deployment pipeline using GitHub Actions, OIDC and MLflow
• ML governance: who promoted which model, trained on what data

Class 6: Model Monitoring + Drift

• Why models fail silently
• Evidently AI for data quality and drift scoring
• Prometheus metrics from inference: confidence, volume, latency
• Automated retraining on drift

Project: A code push triggers training, the challenger is evaluated against the champion, the new model rolls out as an Istio canary, and drift monitoring with retraining runs live.

Module 2: AIOps + Production RAG & AI Platform (5 Weeks)

Week 4: LLM Infrastructure on AWS + MCP

Class 7: LLMs for Platform Engineers

• AIOps defined: where LLMs help operations and where they add noise
• Amazon Bedrock: models, pricing, quotas, IAM
• Local LLM with Ollama for cheap prompt testing
• Prompt engineering for ops: structured outputs, no hallucinated commands

Class 8: MCP: Giving LLMs Access to Your Infrastructure

• The Model Context Protocol and why it matters
• Building a custom MCP server in Python
• Read-only vs write tools, and why that boundary matters

Project: Build an MCP server that answers natural-language questions about AWS cost and resources.

Week 5: Serverless RAG Platform, Part 1: Document Ingestion

Class 9: RAG Architecture on AWS

• What RAG is, and RAG vs fine-tuning
• Serverless RAG architecture: Bedrock, Lambda, Step Functions, S3, DynamoDB
• Bedrock Knowledge Bases vs building your own pipeline, and when each fits
• Embeddings with Bedrock, plus chunking strategies for real documents
• Vector store options: OpenSearch Serverless, pgvector, S3 Vectors

Class 10: Automated Ingestion Pipeline

• S3 upload → EventBridge → Step Functions orchestration
• Lambda steps: parse → chunk → embed → index
• DynamoDB for document metadata, ingestion status and deduplication
• Retries, DLQs and partial failures in Step Functions
• Re-ingestion when a document changes

Project: Build an automated document ingestion pipeline: drop a file in S3 and it is parsed, chunked, embedded and indexed, with status tracked in DynamoDB.

Week 6: Serverless RAG Platform, Part 2: Retrieval, Guardrails + AI Observability

Class 11: Context-Aware Responses + Guardrails

• Retrieval API with API Gateway, Lambda and Bedrock
• Context-aware responses: prompt assembly, citations, conversation history in DynamoDB
• Bedrock Guardrails: PII redaction, denied topics, prompt-injection protection
• Evaluating RAG quality: retrieval precision and answer relevance

Class 12: AI Observability with OpenTelemetry + Langfuse

• Why traditional monitoring misses LLM failures
• OpenTelemetry for GenAI: tracing each retrieval, LLM call and token
• Langfuse self-hosted: traces, prompt versions, cost per request
• Reducing token consumption: smaller chunks, reranking, prompt caching, response caching, model routing
• Measuring improvements in relevance and cost with real numbers, before and after

Project: Build a production RAG API with Guardrails, fully traced in Langfuse. Then optimize it and measure the drop in token cost and the gain in answer relevance.

Week 7: Building AI Agents with Strands Agents

Class 13: Strands Agents Fundamentals

• From chatbot to agent: the reasoning loop and tool calling
• The Strands model-driven approach: a model, a prompt, and tools
• Custom tools with @tool, using the Week 4 MCP server and the Week 6 RAG API as tools
• Bounding the action space: what the agent can and can’t do
• Strands traces flowing into Langfuse

Class 14: Multi-Agent Systems + Deployment

• Multi-agent patterns: agents-as-tools, swarm, graph
• A triage agent delegating to log, metrics and runbook agents
• Deploying on Lambda, ECS Fargate or EKS, and when Bedrock AgentCore makes sense

Project: Build an incident triage agent. An AlertManager webhook triggers it; the agent queries Loki and Prometheus, retrieves runbooks via RAG, and posts severity, probable cause and next actions to Slack.

Week 8: Self-Service AI Platform + Self-Healing Capstone

Class 15: Self-Service AI Infrastructure

• Platform thinking: let teams launch AI workflows without filing tickets
• Reusable Terraform modules: RAG pipeline, agent runtime, MLflow, observability
• GitHub Actions reusable workflows with OIDC: one PR provisions a full AI workflow
• MLflow and Kubernetes integration for automated model tracking, deployment and monitoring
• Observability on by default: every new workflow ships with OpenTelemetry and Langfuse tracing

Class 16: Self-Healing with Guardrails + Capstone

• Detect → decide → act → verify, and where each stage breaks
• Dry-run mode: the agent proposes, a human approves in Slack, then the action runs
• Blast radius limits, max actions per incident, automatic rollback
• Live chaos experiment, with every agent decision traced and reviewed

Capstone: Production RAG & AI Platform. An AWS-based serverless platform for document ingestion, RAG, AI agents, AI observability and reusable MLOps, all delivered through Terraform and GitHub Actions.

What You Will Get

• 8 weeks of live instruction, 16 classes of 3 hours each
• One connected, production-grade AI platform built week by week, ready for your portfolio and resume
• Lifetime access to recordings, code, and notes
• LivingDevOps Discord community and alumni network
• Certificate of completion: MLOps + AIOps