Applied Agentic AI Course

1. Week 1 - LLM Foundations & System Thinking

Understand how LLMs work - Transformer intuition (simplified), Attention mechanism concept, Tokenization & embeddings, Context window limitations, Hallucination causes, RAG failure modes (intro), Prompt vs Retrieval vs Fine-Tuning - overview comparison enough to engineer systems around them.

Topics

• Traditional AI vs Generative AI -paradigm shift explained
• Transformer architecture: attention mechanism, layers, context
• Tokens, embeddings, and context windows - engineering implications
• How LLMs generate text: next-token prediction, temperature, sampling strategies
• LLM limitations: hallucinations, knowledge cutoff, context degradation
• Introduction to information retrieval: why LLMs alone aren't enough
• Overview: OpenAI vs Google Gemini vs Anthropic Claude - capability map
• Safety, responsible AI, and deployment considerations

Hand-ons

• OpenAI Python SDK
• Google Gemini SDK
• Anthropic Claude SDK
• tiktoken
• Visual Studio Code/ Google Colab

Project

Multi-SDK LLM Playground

Build a unified interface that queries OpenAI, Gemini, and Claude simultaneously and compares responses side by side.

2. Week 2 - Prompt Engineering & Structured Output

Master advanced prompting techniques and build systems that produce reliable, structured, evaluable outputs.

Topics

• Zero-shot, few-shot, and chain-of-thought (CoT) prompting - when to use each
• Tree-of-Thought (ToT) and self-consistency prompting
• Role prompting and persona design for consistent behavior
• Designing output constraints: format, length, tone, schema
• Structured responses: JSON schema enforcement, typed outputs
• Function calling and tool use: mechanics and design patterns
• Prompt injection: attack vectors and defense strategies
• Evaluation-driven iteration: test → measure → improve loops
• Introduction to prompt versioning and management

Hand-ons

• OpenAI JSON mode
• Open Ai Playground/Clause Console/Google Ai studio
• LangChain PromptTemplate
• OpenAI Evals (intro)/Claude Console Prompt generator
• PromptLayer

Project

Call Center Transcript Sentiment Classifier

Build a production-grade sentiment + intent classifier for call center transcripts using structured output, function calling, and an evaluation harness to measure accuracy.

3. Week 3 - LLM SDK Mastery & RAG: Foundations to Production

Master SDK capabilities, build a full production RAG pipeline with hybrid search, reranking, RAGAS evaluation, and LangSmith observability.

Topics

• Advanced SDK usage: streaming, async calls, retry logic, rate limiting
• JSON schemas, connectors, and tool calls across OpenAI/Gemini/Anthropic
• Why RAG is essential for enterprise AI — the knowledge freshness problem
• End-to-end RAG pipeline architecture
• Embeddings deep dive: how they work, how to choose the right model
• Introduction to information retrieval: keyword (TF-IDF/BM25) vs semantic search
• Vector databases: internals, HNSW, ANN algorithms
• Chunking strategies: fixed, semantic, recursive, document-aware
• Two-stage retrieval: retrieve-then-rerank with cross-encoders · RAGAS evaluation: faithfulness, relevance, completeness · LangSmith observability and tracing

Hand-ons

• LangChain (RAG chains)
• FAISS / ChromaDB
• OpenAI Embeddings API
• File Search SDK
• RAGAS
• LangSmith / LangFuse

Project

Company Knowledge-Base Chatbot

Complete Build a fully production-grade knowledge-base chatbot: ingest PDFs, websites, and proprietary data, implement hybrid search and reranking, evaluate with RAGAS, and trace with LangSmith

4. Week 4 - Building AI Agents

Build structured AI agents that reason, use tools, maintain memory, and handle failures gracefully.

Topics

• Core components of an AI agent: roles, memory, tools, autonomy, goals
• Agent types: Researcher, Writer, Planner, Analyzer — design patterns
• ReAct framework: Reasoning + Acting in a loop
• Chain-of-Thought vs ReAct vs Tool Use — when to use which
• Tool calling deep dive: function schemas, tool orchestration, tool chaining
• Memory systems: short-term (context), long-term (vector store), episodic
• Connecting agents to APIs, databases, browsers, and file systems
• Memory patterns and structured context passing
• Autonomy levels and safe task-scoping

Hand-ons

• LangChain Agents
• Google ADK
• OpenAI Function Calling
• Anthropic Tool Use
• Mem0 (memory)
• Guardrails AI

Project

Researcher-and-Summarizer Agent + Customer Care Agent

Build two agents: (1) A Researcher-and-Summarizer that searches, reads, and synthesizes information autonomously. (2) A Customer Care Agent that uses tool calling to fetch live order information from.

5. Week 5 - Multi-Agent Systems & Google ADK

Design multi-agent architectures and build production coordinator patterns using CrewAI and Google ADK.

Topics:

• Multi-agent architectures: sequential, parallel, hierarchical patterns
• Role-based agents: specialist vs generalist design
• Delegation logic: how agents assign and route sub-tasks
• Coordinator–dispatcher architecture for production workflows
• Passing context and outputs across agent boundaries
• Background agents vs interactive agents
• File persistence using structured Markdown and JSON reports
• Building custom function tools: financial APIs, file writers, data fetchers
• Designing scalable multi-agent orchestration patterns

Hand-ons

• CrewAI
• Google ADK
• LangGraph (multi-agent)
• Open AI SDK
• OpenAI Assistants API
  • Model Context Protocol (MCP): what it is and why modern agents use it

Project

Financial News Research Coordinator

Build a Financial News Research Coordinator using Google ADK: a coordinator agent dispatches specialized sub-agents (news fetcher, sentiment analyzer, report writer) and produces a structured investment.

Week 6 : MCP, Voice, Multimodal, Evaluation & Capstone

Master MCP architecture, build voice G multimodal agents, evaluate and debug agentic systems, and deliver a fully integrated end-to-end Capstone project.

Evaluating agents end-to-end: task completion rate, tool use accuracy, response quality · Debugging agentic systems: tracing execution and identifying failure points · Reliability patterns: idempotency, retries, fallback strategies · Capstone architecture: RAG + Agents + Multi-Agent + MCP integrated.

Topics Covered:

• MCP architecture: client, server, and transport layer
• Registering tools, services, and capabilities through MCP servers
• How MCP simplifies connecting agents to databases, APIs, and internal systems
• MCP vs traditional tool-calling: comparison and migration path
• Integrating MCP into multi-agent applications
• Gemini multimodal capabilities: vision, audio, documents
• Voice agent architecture: speech-to-text → LLM → action → text-to-speech Hand-ons Practical:
• MCP SDK
• Gemini Multimodal API
• Google Speech-to-Text / TTS
• LangSmith
• Helicone
• RAGAS · WGB · Guardrails AI · FastAPI Project:

MCP-Powered Agent + Voice Assistant + Capstone Build three deliverables:

1. MCP server connecting agents to databases and APIs.
2. Voice-enabled multimodal assistant.
3. CAPSTONE: fully integrated multi-agent application combining RAG, tool- using agents, evaluation loops, and production deployment