Agentic AI Course

1.1 Introduction to AI

• What is Artificial Intelligence?
• AI vs. Automation vs. Analytics: Understanding the distinctions
• Types of AI: Narrow AI, General AI, and Super AI
• Real-world examples of AI in action

1.2 AI Family Tree

• AI, Machine Learning (ML), and Deep Learning (DL): Relationships and differences
• Supervised vs. Unsupervised vs. Reinforcement Learning (brief overview)
• Neural Networks: The building blocks

1.3 History & Evolution of AI

• Key milestones: From Turing Test to AlphaGo to ChatGPT
• AI winters and breakthroughs
• Current state of AI

1.4 Core AI Applications

• Natural Language Processing (NLP): chatbots, translation, sentiment analysis
• Computer Vision (CV): facial recognition, object detection, medical imaging
• Robotics & Autonomous Systems
• Predictive Analytics: forecasting, recommendation systems

1.5 AI Across Industries

• HR: Recruitment screening, employee engagement analysis
• Operations: Supply chain optimization, predictive maintenance
• Healthcare: Diagnosis assistance, drug discovery
• Retail: Personalization, inventory management
• Education: Adaptive learning, automated grading
• Finance: Fraud detection, algorithmic trading
• Manufacturing: Quality control, process automation

1.6 Future of AI

• Emerging trends: Multimodal AI, Edge AI, Quantum ML
• Opportunities: Productivity gains, innovation acceleration
• Risks: Job displacement, privacy concerns, weaponization

2.1 Introduction to Generative AI

• What is Generative AI? Definition and characteristics
• Traditional AI (predictive) vs. Generative AI (creative)
• Everyday GenAI applications: Writing assistants, image generators, code completion
• Popular GenAI tools: ChatGPT, Claude, Gemini, DALL-E, Midjourney

2.2 How GenAI Works

• Large Language Models (LLMs) explained simply
• Introduction to Transformers: Attention mechanism in plain language
• Tokens and tokenization
• Training vs. Fine-tuning vs. Inference
• Parameters and model size (7B, 70B, 405B – what do they mean?)

2.3 Key GenAI Models Overview

• OpenAI GPT family (GPT-4, GPT-4o, o1)
• Anthropic Claude (Sonnet, Opus, Haiku)
• Google Gemini
• Meta LLaMA
• Open-source vs. Closed-source models
• Strengths and limitations of each

2.4 Recent Trends & Important Concepts

• Multimodal AI (text + image + audio + video)
• Context windows and long-form understanding
• Retrieval-Augmented Generation (RAG)
• AI reasoning models (o1, o3)
• Real-time AI and streaming responses

2.5 Prompt Engineering Fundamentals

• What is a prompt?
• Anatomy of a good prompt: Role, Task, Context, Format, Constraints
• Text generation and completion
• Summarization: Short vs. detailed summaries
• Rewriting: Tone adjustment (formal, casual, persuasive, empathetic)
• Translation and localization

2.6 Prompt Structures

• Instructional prompts
• Role-based prompts (“Act as a…”)
• Template-based prompts
• Structured output requests (JSON, tables, lists)

2.7 Core Prompting Techniques

• Zero-shot prompting: Direct instructions
• Few-shot prompting: Learning from examples
• Chain of Thought (CoT): Step-by-step reasoning
• Tree of Thoughts (ToT): Exploring multiple reasoning paths
• Self-consistency: Multiple attempts for better answers

2.8 Multi-turn Interactions

• Context retention and conversation memory
• Building on previous responses
• Clarification and refinement loops

2.9 Limitations & Quality Evaluation

• Hallucinations: What they are and how to spot them
• Bias in AI outputs
• Factual accuracy verification
• Evaluating output quality: Relevance, coherence, accuracy, creativity
• When to use AI vs. human judgment

3.1 LLM Deep Dive

• How transformers process language
• Attention mechanisms visualized
• Pre-training and next-token prediction
• Temperature and sampling parameters
• Top-p, Top-k sampling explained

3.2 LLMs vs. Small Language Models (SLMs)

• When to use LLMs vs. SLMs
• Edge deployment and efficiency
• Cost considerations

3.3 Popular LLM Comparison

• GPT-4: Strengths in reasoning and creativity
• Claude: Long context and nuanced understanding
• Gemini: Multimodal capabilities
• LLaMA: Open-source flexibility
• Use case mapping: Which model for which task?

3.4 Advanced Prompting Techniques

• Self-consistency prompting
• Instruction tuning and prompt optimization
• Negative prompting (what to avoid)
• Constraint-based prompting
• Persona switching and style guides
• Prompt chaining for complex tasks

3.5 Context Management

• Working with large documents
• Context window limitations and strategies
• Summarization for context compression
• Conversation history management

3.6 Introduction to Embeddings

• What are embeddings?
• Vector representations of text
• Semantic search applications
• Use cases: Document similarity, recommendation systems

3.7 Fine-tuning Basics

• What is fine-tuning?
• When to fine-tune vs. prompt engineer
• Transfer learning concept
• No-code fine-tuning platforms (brief overview)

4.1 Introduction to Agentic AI

• What is Agentic AI?
• Key characteristics: Autonomy, goal-directedness, adaptability
• Agentic AI vs. Traditional AI workflows
• Real-world examples: Research agents, customer service agents, coding agents

4.2 AI Agents vs. Chatbots

• Chatbots: Reactive, scripted interactions
• AI Agents: Proactive, goal-oriented, tool-using
• Comparison matrix

4.3 Components of Agentic AI

• Memory: Short-term and long-term context retention
• Planning: Breaking down goals into sub-tasks
• Tool Use: API calls, web search, code execution, database queries
• Autonomy: Self-directed task completion
• Reflection: Self-evaluation and improvement

4.4 Agent Frameworks Overview

• LangChain: Modular components for agent building
• AutoGen: Multi-agent conversations
• CrewAI: Role-based agent collaboration
• OpenAI Assistants API: Built-in agent capabilities
• Comparison and use case mapping

4.5 AI Orchestration

• What is orchestration?
• Workflow design for agent tasks
• Human-in-the-loop (HITL) integration
• Error handling and fallback strategies

5.1 No-Code AI Overview

• What is no-code/low-code?
• Benefits: Speed, accessibility, cost-effectiveness
• Limitations: Customization constraints, vendor lock-in
• When to use no-code vs. custom development

5.2 Workflow Automation Fundamentals

• Triggers, actions, and conditions
• API integrations basics
• Data mapping and transformation
• Error handling and testing

5.3 Zapier

• Platform overview and interface
• Building your first Zap
• Multi-step workflows
• AI-powered Zaps with ChatGPT integration
• Filters, formatters, and utilities
• Best practices and optimization
• Hands-on: Automate email-to-task workflow

5.4 Make.com

• Visual workflow builder
• Modules, routes, and scenarios
• Advanced routing and error handling
• Data stores and aggregators
• Scheduling and webhooks
• Hands-on: Build a content aggregation workflow

5.5 n8n

• Self-hosted automation (overview)
• Node-based workflow design
• Custom code nodes
• Integrations and credentials management
• Hands-on: Create a Slack notification system

5.6 Notion AI

• AI writing and editing
• Database automation
• Template creation with AI
• Q&A over workspace knowledge
• Hands-on: Build an AI-powered knowledge base

5.7 Airtable

• AI-powered data management
• Automation with AI fields
• Integration with other tools
• Building mini-apps
• Hands-on: Create a project tracker with AI summaries

5.8 Glide

• No-code app builder
• AI-powered features
• Data binding and workflows
• Mobile app prototyping
• Hands-on: Build a simple internal tool

5.9 Canva AI

• Text-to-design generation
• AI brand kit creation
• Magic Resize and Magic Eraser
• AI-powered content suggestions
• Hands-on: Create a presentation with AI

5.10 Tome AI

• AI-powered storytelling
• Auto-generating narratives
• Interactive presentations
• Hands-on: Build a pitch deck

5.11 Gamma.app

• AI slide deck creation
• One-click formatting
• Collaborative editing
• Hands-on: Generate a training module

5.12 Integrating AI into Business Processes

• Identifying automation opportunities
• Workflow mapping and optimization
• Change management considerations
• Measuring ROI of automation
• Building a toolkit for your role

6.1 Agent Design Principles

• Defining agent goals and scope
• Task decomposition
• User experience considerations
• Edge case handling

6.2 No-Code Agent Platforms

• Relevance AI: Building AI teams
• Stack AI: Workflow-based agents
• Flowise: Visual LLM app builder
• Voiceflow: Conversational agents
• Platform comparison and selection

6.3 Building Your First Agent

• Step-by-step agent creation
• Defining instructions and personality
• Adding tools and integrations
• Knowledge base integration (RAG)
• Testing and iteration

6.4 Practical: AI Research Assistant Agent

• Goal: Automated web research and summarization
• Components:
  1. Web search capability
  2. Content extraction
  3. Summarization
  4. Report generation
  5. Email delivery
• Tools: n8n/Make + GPT API + Google Docs
• Build process:
  1. Design the workflow
  2. Set up web search module
  3. Configure summarization
  4. Create output template
  5. Add scheduling
  6. Test and refine

6.5 Advanced Agent Features

• Memory and context persistence
• Multi-tool orchestration
• Conditional logic and branching
• User input handling
• Feedback loops

6.6 Deployment & Monitoring

• Publishing your agent
• Usage analytics
• Performance monitoring
• Iterative improvement
• Cost management

7.1 AI Challenges

→ Bias

• What is AI bias?
• Sources: Training data, algorithm design, human feedback
• Types: Gender, racial, cultural, socioeconomic
• Real-world examples and consequences
• Detection and mitigation strategies

→ Hallucinations

• What are hallucinations?
• Why LLMs hallucinate
• Identifying hallucinations
• Mitigation: Verification, grounding, confidence scoring

→ Privacy & Security

• Data privacy concerns
• Prompt injection attacks
• Data leakage risks
• Secure AI usage practices
• Compliance considerations (GDPR, CCPA)

→ Misinformation & Deepfakes

• AI-generated content detection
• Deepfake risks
• Watermarking and provenance
• Media literacy in AI age

7.2 Responsible AI Principles

• Fairness: Eliminating discriminatory outcomes
• Accountability: Clear ownership and responsibility
• Transparency: Explainability and disclosure
• Explainability: Understanding AI decisions
• Privacy: Data protection and consent
• Safety & Security: Robustness and resilience
• Human Control: Human-in-the-loop systems

7.3 Global AI Ethics Frameworks

→ OECD AI Principles

• Inclusive growth and sustainability
• Human-centered values
• Transparency and explainability
• Robustness and safety
• Accountability

→ EU AI Act

• Risk-based approach
• Prohibited AI practices
• High-risk AI systems
• Transparency obligations
• Compliance requirements

→ NITI Aayog (India)

• #AIForAll vision
• Responsible AI strategy
• Focus on explainability and fairness
• Sector-specific guidelines

→ UNESCO Recommendation on AI Ethics

• Human rights and dignity
• Environmental sustainability
• Cultural diversity
• Gender equality

→ Microsoft Responsible AI

• Six principles framework
• Impact assessments
• Governance structure
• Practical tools

7.4 Safe & Responsible Usage

→ Verifying AI Outputs

• Cross-referencing with trusted sources
• Fact-checking methodologies
• Using multiple AI models for comparison
• Critical evaluation frameworks

→ Human-in-the-Loop (HITL) Systems

• When to require human oversight
• Designing HITL workflows
• Balancing automation and control
• Decision authority frameworks

→ Trust & Compliance

• Building organizational AI policies
• Training and awareness programs
• Audit trails and documentation
• Incident response plans

7.5 Responsible AI Guardrails

• Input validation and filtering
• Output moderation
• Rate limiting and abuse prevention
• Content policies and guidelines
• Monitoring and alerting systems

7.6 Case Studies of AI Misuse

• Amazon’s biased recruiting tool
• COMPAS recidivism algorithm
• Deepfake political videos
• ChatGPT jailbreaks and prompt injection
• Lessons learned and best practices

7.7 Building Ethical AI Culture

• Organizational responsibilities
• Individual accountabilities
• Ethical decision-making frameworks
• Continuous learning and adaptation

1.1 Development Environment Setup

• Installing and configuring VS Code for AI development
• Jupyter Lab setup and best practices
• GitHub fundamentals (clone, commit, push, pull requests)
• Virtual environments (venv, conda)
• Best practices for managing dependencies (requirements.txt, poetry, pip-tools)
• Workspace optimization and productivity tips

1.2 Python Essentials for AI

• Strings and text processing for NLP
• Iterators and generators (memory-efficient data handling)
• List comprehensions and functional programming
• Threads and processes for parallel AI workflows
• Asynchronous programming basics (async/await)

1.3 File Operations & Web Crawling

• Working with CSV files (reading, writing, processing large files)
• JSON handling (parsing, serialization, nested structures)
• PDF extraction (PyPDF2, pdfplumber)
• Excel files (openpyxl, pandas)
• Web scraping basics (BeautifulSoup, requests)
• Crawling best practices and ethics

1.4 NumPy & Pandas for AI

• NumPy arrays and vectorized operations
• Broadcasting and advanced indexing
• Pandas DataFrames for data manipulation
• Data cleaning and preprocessing
• Grouping, aggregation, and transformation
• Time series handling
• Integration with AI workflows

1.5 Functions & Classes

• Function design patterns for AI
• Decorators and context managers
• Object-oriented programming for AI projects
• Class design for agents and workflows
• Inheritance and composition
• Type hints and documentation

1.6 Working with APIs

• REST API fundamentals
• requests library deep dive
• Authentication (API keys, OAuth)
• JSON handling and parsing
• Error handling and retries
• Rate limiting and pagination
• Building API wrappers

2.1 Transformers Architecture

• Attention mechanism explained visually
• Self-attention and multi-head attention
• Encoder-decoder architecture
• Positional encodings
• Layer normalization and residual connections
• BERT (encoder-only) vs GPT (decoder-only) vs T5 (encoder-decoder)

2.2 Hugging Face Ecosystem

• Hugging Face Transformers library overview
• Model Hub exploration
• Tokenizers (WordPiece, BPE, SentencePiece)
• Pipeline API for quick inference
• AutoModel, AutoTokenizer classes
• Model cards and documentation

2.3 Pre-trained vs Fine-tuned Models

• Transfer learning fundamentals
• When to use pre-trained vs fine-tuned models
• Understanding model capabilities and limitations
• Popular models: BERT, RoBERTa, GPT-2, T5, BART
• Model selection criteria for business use cases

2.4 Fine-Tuning Pre-trained Models

• Fine-tuning BERT for classification
• Fine-tuning GPT for text generation
• Dataset preparation and formatting
• Training arguments and hyperparameters
• Trainer API and training loops
• Evaluation metrics and validation
• Saving and loading fine-tuned models

2.5 Parameter-Efficient Fine-Tuning (PEFT)

• Why PEFT? (cost, speed, memory efficiency)
• Low-Rank Adaptation (LoRA) explained
• QLoRA (Quantized LoRA)
• Adapter layers
• Prefix tuning and prompt tuning
• Comparing PEFT methods
• Implementation with PEFT library

2.6 Advanced Fine-Tuning Techniques

• Instruction tuning
• Reinforcement Learning from Human Feedback (RLHF) overview
• Multi-task fine-tuning
• Domain adaptation
• Catastrophic forgetting mitigation

Hands-on 

• Load and test pre-trained models from Hugging Face
• Fine-tune BERT for sentiment analysis on custom dataset
• Fine-tune GPT-2 for domain-specific text generation
• Implement LoRA fine-tuning for large model on limited GPU

3.1 Language Modeling Fundamentals

• What is language modeling?
• Next-token prediction
• Probability distributions over vocabulary
• Autoregressive models
• Causal vs masked language modeling

3.2 Tokenization & Embeddings

• Tokenization strategies (character, word, subword)
• Building custom tokenizers
• Token embeddings and position embeddings
• Embedding dimensions and vocabulary size
• Learned vs fixed positional encodings

3.3 Building GPT Step-by-Step

• Implementing multi-head self-attention
• Building transformer blocks
• Stacking layers
• Adding feedforward networks
• Layer normalization placement
• Residual connections
• Output projection layer

3.4 Training Pipeline

• Preparing training data
• Creating data loaders
• Implementing training loop
• Loss function (cross-entropy)
• Optimizer selection (Adam, AdamW)
• Learning rate scheduling
• Gradient clipping
• Checkpointing

3.5 Sampling & Generation

• Greedy decoding
• Beam search
• Top-k and top-p (nucleus) sampling
• Temperature scaling
• Repetition penalty
• Length constraints

Hands-on

• Implement tokenizer and embeddings from scratch
• Build complete GPT architecture in PyTorch
• Train mini-GPT on toy dataset and generate text

4.1 Introduction to Multimodal AI

• What is multimodal AI?
• Why combine modalities?
• Challenges in multimodal learning
• Business applications across industries
• State-of-the-art multimodal models

4.2 Multimodal Data Representation

• Datatype comparisons (text, image, audio, video)
• Metadata handling and alignment
• Representation learning across modalities
• Feature extraction strategies
• Normalization and standardization

4.3 Individual Modality Processing

• Text Processing: Tokenization, Normalization, Stopword removal, Stemming/lemmatization, Named entity recognition
• Image Processing: Resizing, Normalization, Augmentation, Feature extraction with CNNs, Image embeddings
• Audio Processing: Sampling rates, Spectrograms, MFCCs (Mel-frequency cepstral coefficients), Audio augmentation, Speech features
• Video Processing: Frame extraction, Temporal sampling, Optical flow, Video embeddings
• Preprocessing Challenges: Missing data, Noise, Format inconsistencies, Computational costs

4.4 Text-to-Image Models

• Stable Diffusion architecture and usage
• DALL·E 3 API integration
• Diffusion models explained
• Prompt engineering for image generation
• ControlNet and IP-Adapter
• Fine-tuning diffusion models (DreamBooth, LoRA)

4.5 Image-to-Text Models

• CLIP (Contrastive Language-Image Pre-training)
• BLIP (Bootstrapping Language-Image Pre-training)
• Flamingo and few-shot learning
• Image captioning
• Visual question answering (VQA)
• Document understanding (LayoutLM, Donut)

4.6 Speech Processing

• Text-to-Speech: ElevenLabs, OpenAI TTS API, Coqui TTS, Voice cloning
• Speech-to-Text: OpenAI Whisper (local and API), Google Speech-to-Text, Assembly AI, Real-time transcription, Speaker diarization, Multilingual support

4.7 Combining Modalities

• Vision + Text pipelines (image search, visual QA)
• Text + Audio (podcast summarization, meeting transcription)
• Vision + Audio (video understanding)
• Multi-modal embeddings and alignment
• Fusion strategies (early, late, hybrid)

4.8 Popular Multimodal Datasets

• COCO (Common Objects in Context)
• Visual Genome
• Conceptual Captions
• LAION
• AudioSet
• VoxCeleb
• HowTo100M
• Dataset exploration and statistics
• Efficient data loading strategies
• Custom dataset creation

4.9 Data Alignment and Synchronization

• Temporal alignment across modalities
• Cross-modal validation
• Handling missing data
• Annotation and labeling
• Building alignment pipelines
• Quality assurance

4.10 Evaluation Metrics and Benchmarks

• Vision-Language Metrics: BLEU, METEOR, ROUGE, CIDEr for captioning
• Semantic Evaluation: CLIP score, Semantic similarity
• Audio Metrics: WER (Word Error Rate), Speaker recognition accuracy
• Multimodal Benchmarks: VQA, GQA, NLVR2, TVQA

Hands-on 

• Generate images with Stable Diffusion and prompt engineering
• Build image captioning system with BLIP
• Implement speech-to-text with Whisper (transcription + summarization)
• Create text-to-speech voice assistant
• Build multimodal search (text query → image results)
• Video analysis pipeline (extract frames → caption → summarize)

5.1 LLM Wrappers & APIs

• OpenAI API (GPT-4, GPT-4o)
• Anthropic Claude API
• Google Gemini API
• Hugging Face Inference API
• Ollama for local models
• API authentication and rate limits
• Cost optimization strategies

5.2 Advanced Prompt Engineering

• System prompts and role definition
• Few-shot learning with examples
• Chain of Thought (CoT) prompting
• Tree of Thoughts (ToT)
• Self-consistency
• ReAct (Reasoning + Acting)
• Prompt templates and versioning

5.3 LangChain Fundamentals

• LangChain architecture overview
• Core components: Models, Prompts, Chains, Agents, Memory
• LLM wrappers (OpenAI, Anthropic, HuggingFace)
• PromptTemplates and few-shot examples
• Output parsers (JSON, structured data)
• LangChain Expression Language (LCEL)
• Runnables and parallel execution

5.4 LangChain Memory

• ConversationBufferMemory
• ConversationSummaryMemory
• ConversationBufferWindowMemory
• VectorStoreMemory
• Entity memory
• Custom memory implementations

5.5 LangChain Chains

• LLMChain basics
• Sequential chains
• Router chains
• Transform chains
• MapReduce for document processing
• Custom chain creation
• Chain debugging

5.6 LangChain Tools

• Built-in tools (Calculator, Wikipedia, Google Search, Python REPL)
• Custom tool creation
• Tool selection and routing
• Error handling in tools
• API-based tools

5.7 LangChain Agents

• Agent types (Zero-shot ReAct, Conversational, OpenAI Functions)
• Agent executors
• Multi-action agents
• Agent memory and state
• Building custom agents
• Agent orchestration

5.8 Production Considerations

• Streaming responses
• Caching for performance
• Rate limiting and retry logic
• Error handling and fallbacks
• Logging and observability
• Token counting and cost tracking

Hands-on 

• Build conversational chatbot with memory using LangChain
• Create document Q&A chain with MapReduce
• Implement ReAct agent with custom tools
• Build research assistant with web search and summarization
• Create multi-step workflow with sequential chains
• Build visual workflow with LangFlow (chatbot or RAG system)

6.1 Introduction to Embeddings

• What are embeddings?
• Vector representations of text
• Semantic similarity
• Embedding dimensions
• Use cases for embeddings

6.2 Embedding Models Comparison

• OpenAI Embeddings
• Sentence Transformers
• E5 and BGE models
• Cohere Embeddings
• Sparse Models
• Comparison criteria: Accuracy, Speed, Cost, Dimensionality, Multilingual support

6.3 Vector Similarity Search

• Cosine similarity
• Dot product
• Euclidean distance
• Approximate nearest neighbors (ANN)
• Index types (flat, IVF, HNSW)
• Trade-offs: speed vs accuracy

6.4 Vector Databases

• FAISS: Local vector search, Index types, GPU acceleration
• Pinecone: Managed cloud service, Namespaces, Metadata filtering

6.5 Document Processing & Chunking

• Text splitting strategies
• Chunk size optimization
• Overlap considerations
• Semantic chunking: Splitting by meaning rather than length
• Content-aware chunking: Preserving document structure (headers, paragraphs)
• Markdown-aware splitting
• Code-aware splitting
• Metadata preservation

6.6 Indexing Strategies

• Document loaders (PDF, DOCX, HTML, Markdown)
• Text preprocessing
• Embedding generation in batches
• Upserting to vector databases
• Metadata tagging
• Version control for knowledge bases

6.7 Building RAG Pipelines

• RAG architecture overview
• Query → Retrieval → Augmentation → Generation
• Context window management
• Prompt engineering for RAG
• Citation and source attribution
• Handling multi-document context

6.8 Advanced Retrieval Techniques

• Hybrid search: Combining dense (vector) and sparse (keyword) search
• Metadata filtering: Pre-filtering and post-filtering
• Multi-query retrieval: Generating multiple search queries
• Parent document retrieval: Retrieving full context
• Ensemble retrieval: Combining multiple retrievers

6.9 RAG Fusion & Re-ranking

• • RAG Fusion: Reciprocal Rank Fusion (RRF) for combining results
  • Re-ranking models: Cohere Rerank, Cross-encoders, FlagEmbedding
  • Two-stage retrieval: Fast retrieval + accurate re-ranking
  • Diversity and relevance balancing

6.10 Building Production RAG Chatbot

• Conversational RAG with memory
• Context compression
• Streaming responses
• Chat history management
• Source citation in UI
• Evaluation and testing

6.11 RAG Evaluation

• Retrieval metrics (precision, recall, MRR, NDCG)
• Generation quality (faithfulness, relevance)
• End-to-end evaluation frameworks
• A/B testing retrieval strategies

Hands-on

• Generate and compare embeddings from different models
• Build vector search system with FAISS
• Implement semantic chunking for complex documents
• Create basic RAG pipeline with LangChain + Pinecone
• Implement hybrid search with re-ranking

7.1 Introduction to Autonomous Agents

• What makes AI “agentic”?
• Agent vs chatbot vs assistant
• Key components: Memory, Planning, Tools, Reasoning, Action
• Agent architectures (ReAct, Plan-and-Execute, Reflexion)
• Real-world agent use cases

7.2 Agent Frameworks Overview

• • LangGraph: State machines for agents
  • CrewAI: Role-based multi-agent teams
  • AutoGen: Conversational agent framework
  • AgentGPT: Autonomous task completion
  • Agno-Light Framework: Production-ready agent framework, Built-in tools and integrations, Workflow orchestration
  • LangFlow: Visual workflow builder for LangChain,  Drag-and-drop interface, Low-code agent development
  • Framework comparison and selection

7.3 CrewAI Deep Dive

• Agents, Tasks, and Crews
• Defining agent roles and goals
• Tools and capabilities
• Process types (sequential, hierarchical)
• Inter-agent communication
• Task delegation and execution
• Memory and context sharing

7.4 AutoGen Framework

• Conversational agents
• AssistantAgent and UserProxyAgent
• GroupChat for multi-agent conversations
• Code execution agents
• Human-in-the-loop workflows
• Nested chats and handoffs

7.5 Agno Framework

• Introduction to agno architecture
• Building production-ready agents
• Structured outputs and type safety
• Built-in tools (web search, file operations, API calls)
• Knowledge bases and memory
• Agent teams and orchestration
• Storage and state management
• Monitoring and debugging agents

7.6 LangFlow for Visual Agent Development

• Introduction to LangFlow interface
• Drag-and-drop workflow design
• Components: LLMs, Chains, Agents, Tools, Memory
• Building chatbots visually
• Creating RAG systems without code
• Custom components and Python integration
• Exporting workflows to code
• Deployment from LangFlow
• LangFlow vs traditional coding approach

7.7 Task Decomposition & Planning

• Breaking complex goals into subtasks
• Dependency mapping
• Dynamic planning and replanning
• Handling failures and retries
• Backtracking and alternative paths

7.8 Agent Tools & Capabilities

• Web search and browsing
• Code execution (Python, SQL)
• API interactions
• File operations
• Database queries
• Custom tool creation
• Tool selection strategies

7.9 Multi-Agent Coordination

• Agent communication protocols
• Role assignment and specialization
• Consensus mechanisms
• Conflict resolution
• Load balancing
• Asynchronous agent execution

7.10 Memory & State Management

• Short-term memory (working memory)
• Long-term memory (episodic memory)
• Shared memory across agents
• State persistence
• Knowledge graphs for agents

7.11 Reasoning & Decision Making

• Chain of Thought reasoning
• Self-reflection and critique
• Uncertainty handling
• Confidence scoring
• Explainable agent decisions

Hands-on

• Build single autonomous agent with planning and tools (LangChain)
• Create multi-agent research team with CrewAI
• Implement AutoGen conversation between specialist agents
• Build production agent with agno framework
• Create visual RAG chatbot with LangFlow (no-code)

8.1 Introduction to LangGraph

• Why LangGraph? (limitations of simple chains)

• Graph-based workflow design

• Nodes, edges, and state

• Conditional routing

• Cycles and loops

• Checkpointing and persistence

8.2 Building Workflows with LangGraph

• Defining state schema

• Creating nodes (functions)

• Adding edges (transitions)

• Conditional edges

• Entry and exit points

• Compiling graphs

8.3 State Management

• Centralized state vs distributed state

• State updates and merges

• Reducer functions

• State persistence

• Rollback and versioning

• Debugging state transitions

8.4 Advanced LangGraph Patterns

• Human-in-the-loop interrupts

• Sub-graphs and composition

• Error handling and fallbacks

• Parallel execution

• Dynamic graph construction

• Streaming outputs from graphs

8.5 Model Context Protocol (MCP)

• What is MCP?

• Standardized context sharing across tools and models

• MCP servers and clients

• Building MCP-compatible agents

• Context serialization and deserialization

• Future of interoperable AI ecosystems

8.6 Agent-to-Agent Protocols (A2A)

• Inter-agent communication standards

• Message passing patterns

• Discovery and registration

• Authentication and authorization

• Protocol versioning

• Building A2A-compatible systems

8.7 Production Agent Systems

• Agent orchestration platforms

• Monitoring and observability

• Performance optimization

• Cost management

• Security considerations

• Scaling agent systems

8.8 Future of Agentic AI

• Emerging trends (autonomous research, code generation, creative agents)

• Multi-modal agents

• Embodied AI

• Ethical considerations

• Regulatory landscape

Hands-on

• Build simple workflow with LangGraph (routing and conditions)
• Create complex multi-step agent with state persistence
• Implement human-in-the-loop approval workflow
• Design and build production agent system with monitoring

9.1 Testing & Evaluating LLM Applications

• Output Quality Metrics: Relevance and coherence, Factual accuracy, Completeness, Consistency
• Evaluation Frameworks: RAGAS, TruLens, DeepEval, LangSmith Evaluators
• Human evaluation: Creating rubrics, A/B testing
• Automated evaluation: LLM-as-judge, Reference-based metrics

9.2 Hallucination Detection & Mitigation

• What are hallucinations?
• Detection techniques (consistency checking, source attribution, confidence scoring)
• Mitigation strategies (RAG, fact-checking tools, human verification)
• Hallucination benchmarks

9.3 LLM Benchmarking

• Running benchmarks
• Interpreting benchmark results: Trade-offs between capability areas
• Custom benchmarks for your use case

9.4 Embedding Benchmarking

• Retrieval Benchmarks
• Evaluation metrics: Recall@k, MRR, NDCG
• Domain adaptation evaluation: Testing embeddings on your data
• Choosing the best embedding model: Accuracy vs speed vs cost

9.5 Model Selection & Optimization

• Use case-based selection: Capabilities needed vs cost constraints
• Latency requirements: Real-time vs batch processing
• Cost-performance trade-offs: GPT-4 vs GPT-3.5 vs open-source
• Custom benchmarking workflow
• When to use smaller models: Task-specific fine-tuning vs larger general models

9.6 Monitoring & Debugging with LangSmith

• LangSmith setup and configuration
• Tracing LLM calls
• Debugging chains and agents
• Performance analytics
• Cost tracking
• Dataset creation
• Online evaluation
• Feedback collection

9.7 Logging & Observability

• Structured logging best practices
• Token usage tracking
• Latency monitoring
• Error rate tracking
• User feedback loops
• Analytics dashboards

9.8 Application Deployment

• Streamlit: Rapid prototyping, Interactive UIs, Deployment to Streamlit Cloud
• Gradio: Quick demos, Shareable links, Hugging Face Spaces integration
• When to use Streamlit vs Gradio

9.9 API Deployment with FastAPI

• Building REST APIs for LLM applications
• Asynchronous endpoints
• Request/response models with Pydantic
• Authentication and authorization
• Rate limiting
• CORS configuration
• API documentation (OpenAPI/Swagger)
• Containerization with Docker
• Deployment to cloud (AWS, GCP, Azure)

9.10 Production Best Practices

• Environment management (dev, staging, prod)
• Configuration management
• Secrets and API key handling
• Caching strategies (Redis, in-memory)
• Load balancing
• Auto-scaling
• CI/CD pipelines
• Monitoring and alerting
• Incident response
• Cost optimization

9.11 Security & Compliance

• Input validation and sanitization
• Prompt injection prevention
• Data privacy (PII handling, GDPR)
• Model access controls
• Audit logging
• Compliance frameworks

→ Hands-on 

• Set up LangSmith and trace LLM application
• Run benchmark comparison of 3 LLMs on custom test set
• Build and deploy Streamlit app for RAG chatbot
• Create FastAPI endpoint for LLM service
• Implement monitoring dashboard with logging and metrics
• Containerize and deploy application to cloud