← Back to Training Material
Reference Curriculum

Generative AI & Agentic AI — A Hands-on Journey

An intensive, 5-day hands-on reference curriculum exploring the foundations of large language models, application building using free and open SDKs, multi-agent frameworks, and academic research acceleration.

Designed By Akshar Prabhu Desai
Duration 5 Full Days
Daily Timing 10:00 AM – 5:00 PM
Mode In-person, Hands-on Lab
Important Notice: We do not run these bootcamps ourselves. This syllabus is provided strictly as an ideal reference structure designed by experts. Individual trainers, educators, and institutions are encouraged to adopt, modify, and utilize this curriculum for their own training programs.

Program Objectives

By the end of this 5-day training, participants will be able to:

Target Audience

This program is designed for faculty members, researchers, students, and industry professionals who wish to transition from theoretical understanding to building live, production-ready AI applications.

Prerequisites

Basic familiarity with the Python programming language (loops, functions, and elementary libraries) and a laptop with stable internet access are required for hands-on sessions.

Daily Time Structure

Slot Time Duration
Session 1 10:00 AM – 11:15 AM 75 min
Tea Break 11:15 AM – 11:30 AM 15 min
Session 2 11:30 AM – 1:00 PM 75 min
Lunch Break 1:00 PM – 2:00 PM 60 min
Session 3 2:00 PM – 3:15 PM 75 min
Tea Break 3:15 PM – 3:30 PM 15 min
Session 4 3:30 PM – 5:00 PM 75 min

Pre-Training Setup Guide

Participants must complete the following setup instructions before Day 1. A pre-training virtual orientation (1 hour) will be scheduled to help verify developer environments.

1. Accounts to Create (All Free)

Account Purpose Sign Up URL
Google Account Colab, NotebookLM, Google AI Studio (Gemini API) accounts.google.com
Hugging Face Free models, datasets, Spaces, Inference API huggingface.co
Overleaf LaTeX paper writing (free tier) overleaf.com
GitHub Source control, free code hosting github.com
Streamlit Community Cloud Free app deployment platform streamlit.io/cloud
GitHub Copilot Optional, for code suggestions in VS Code github.com/features/copilot

2. Local Software Installation

  • Python 3.10+ — Install Python from the official downloads page: python.org/downloads/
  • VS Code — IDE of choice for coding: code.visualstudio.com/
  • Git — Version control tool: git-scm.com/
  • Ollama — Running local models. Pull lightweight models after install: ollama pull llama3.2:3b && ollama pull phi3:mini
  • Python packages — Install these libraries in a fresh virtual environment (venv): 
    pip install jupyter pandas numpy matplotlib scikit-learn \
            langchain langchain-google-genai langgraph \
            crewai pyautogen \
            chromadb streamlit gradio \
            python-dotenv requests

3. Web Tools to Bookmark

Web interfaces:
Developer Portals:
Research Systems:

Day-Wise Detailed Agenda

Click on each day to expand the detailed sessions, labs, and themes.

Day 1 — Foundations of Generative AI Theme: Setup, Python warm-up, transformer intuition, prompt engineering. (Guided)

Session 1 — Welcome, Setup Verification & Python Warm-up

  • Welcome address, training overview, course expectations.
  • Environment verification (Colab login, Ollama running, package dependencies verified).
  • Python warm-up in Google Colab:
    • NumPy & pandas essentials for working with tensors and dataframes.
    • Vectors, matrices, dot products — the mathematical groundwork behind embeddings.
    • Data visualization and plotting using matplotlib.
    • Quick scikit-learn refresher (train/test splits, running basic classifiers).

Session 2 — The AI/ML/DL/GenAI Landscape

  • AI vs. ML vs. Deep Learning vs. Generative AI: clear structural boundaries.
  • Discriminative models vs. generative models.
  • A brief history of sequence modeling: from RNNs and LSTMs to the modern Transformer.
  • Survey of contemporary generative modalities (text, image, audio, code, video).
  • Cross-model demonstration: comparing ChatGPT, Claude, and Gemini on identical prompts.

Session 3 — Transformer Architecture Deep-Dive

  • Tokens, embeddings, and positional encoding mechanisms.
  • Self-attention and multi-head attention: an intuitive mathematical explanation.
  • Architecture types: Encoder-only (BERT), decoder-only (GPT), and encoder-decoder (T5).
  • Live demo: Transformer Visualizer (calm-sand-0f2fbf80f.7.azurestaticapps.net):
    • Visual walkthrough of attention weights and layer-by-layer tensor flows.
    • Hands-on exploration of visualizer states by participants on their laptops.
  • How GPT-style models generate text: next-token prediction, sampling methods, temperature, and top-p.

Session 4 — Prompt Engineering Hands-on

  • Anatomy of a production prompt: assigning roles, defining context, tasks, formatting, and constraints.
  • Patterns: zero-shot, few-shot, chain-of-thought, role prompting, and the ReAct (Reasoning + Acting) loop.
  • Hands-on exercises in free web UIs (ChatGPT, Claude, Gemini):
    • Summarization, text classification, semantic extraction, reasoning, and code generation.
  • Prompt comparison lab: running matching prompts across 3 models to analyze differences.
  • Daily reflection (10 min quiz/feedback) & preview of Day 2 topics.
Day 2 — LLMs, APIs & Application Building Theme: From understanding LLMs to building real applications. (Guided)

Session 1 — Inside an LLM

  • Tokenization in depth (BPE, WordPiece) — hands-on usage with Hugging Face tokenizers library.
  • Embeddings: mapping words, sentences, and full documents to vector space.
  • Hands-on script: generating sentence embeddings with sentence-transformers and visualizing clusters using PCA/t-SNE.
  • Pre-training vs. fine-tuning vs. RLHF (conceptual structures).
  • Context windows, latency issues, and API pricing/cost considerations.

Session 2 — Hugging Face Ecosystem

  • Hugging Face Hub tour: searching models, datasets, Spaces, and analyzing leaderboards.
  • Loading and running lightweight models locally using the transformers library.
  • Hugging Face Inference API (free tier) — making calls to hosted models from Python scripts.
  • Mini-lab: building sentiment analysis, text classification, summarization, and translation pipelines.

Session 3 — LangChain with Free LLMs

  • LangChain core orchestration concepts: models, prompts templates, output parsers, and chains.
  • Google AI Studio (Gemini API): Generate a free API key and establish your first API call.
  • Ollama: Interfacing with local Llama 3.2 / Phi-3 models from LangChain code.
  • Building code to switch seamlessly between cloud APIs (Gemini) and local endpoints (Ollama).
  • Structuring model outputs using Pydantic schemas.

Session 4 — Build a Chatbot

  • Maintaining conversation memory in LangChain.
  • Building a sleek interactive chatbot user interface in **Streamlit**.
  • Adding system instructions, custom personas, and guardrails to filter inputs/outputs.
  • Deploying the chatbot application on local developer machines.
  • Daily reflection (10 min quiz/feedback) & preview of Day 3 topics.
Day 3 — Agentic AI: Concepts & Frameworks Theme: Beyond chat — agents that plan, use tools, and reason. (Semi-guided)

Session 1 — What is Agentic AI?

  • Moving beyond chat interfaces: planning, reasoning, tool-use, stateful memory, and self-reflection.
  • Anatomy of a standalone agent: LLM brain + tool definitions + memory states + control loop.
  • ReAct (Reason + Action) loop pattern explained step-by-step.
  • Comparing architectural paradigms: Single-agent vs. multi-agent systems.
  • When NOT to use agents (cost structures, accuracy thresholds, latency limitations).

Session 2 — Agent Frameworks Tour

  • LangGraph: Graph-based state orchestration and cycle management (deep code walkthrough).
  • CrewAI: Role-based collaborative multi-agent architecture.
  • AutoGen: Conversational, event-driven multi-agent framework patterns.
  • Comparison framework: selecting the correct orchestrator for specific problem spaces.
  • Hands-on: building the same execution task in two different frameworks to compare design.

Session 3 — Hands-on: Build a Research Agent

  • Defining agent scope: "Given a research topic, query for papers, scrape details, and output summaries."
  • Integrating API tools: DuckDuckGo search library, Wikipedia API, and arXiv academic query.
  • Constructing the agent graph using LangGraph & free Gemini API keys.
  • Iterative refinement: adding persistence memory, error handling boundaries, and retry limits.

Session 4 — Agent Design Patterns & Evaluation

  • Design patterns: planner-executor, agentic reflection, human-in-the-loop (HITL), and supervisor control.
  • Evaluating agent behaviors: execution trajectory quality, tool invocation accuracy, final answers verification.
  • Hands-on: adding logging hooks to trace and evaluate the research agent's steps.
  • Daily reflection (10 min quiz/feedback) & preview of Day 4 topics.
Day 4 — AI for Research & Academia Theme: Practical AI workflows for researchers and academic authors. (Semi-guided)

Session 1 — NotebookLM for Research

  • What NotebookLM is and why it represents a paradigm shift for academic study.
  • Importing source material: PDFs, web URLs, Google Docs, and YouTube transcripts.
  • Synthesizing questions with source citation (eliminating hallucinations).
  • Generating structured study guides, briefings, FAQs, and conceptual mind maps.
  • Audio Overviews: Creating interactive podcast-style audio summaries from paper collections.
  • Hands-on: building a personalized notebook index around 5 academic papers in participants' fields.
  • Best practices: recognizing source quality constraints and boundary thresholds.

Session 2 — Writing Papers with Overleaf + AI

  • Overleaf foundation: setting up projects, editing templates, compiler errors, and collaboration options.
  • Selecting appropriate academic templates (IEEE, ACM, Springer).
  • LaTeX syntax essential overview: headings, math equations, tables, figures, and BibTeX citations.
  • AI-Assisted LaTeX Workflow:
    • Generating structural outlines with LLMs.
    • Paragraph editing for academic tone, clarity, and conciseness.
    • Grammar checking and style adjustments.
    • Converting DOIs directly into BibTeX references.
    • Translating descriptive text into clean LaTeX math equations.
  • Ethics: AI disclosures, plagiarism checks, journal publisher policies on LLM usage.
  • Hands-on: drafting a 1-page paper stub using the integrated workflow.

Session 3 — RAG over Personal Research Corpus

  • RAG (Retrieval-Augmented Generation) concepts: Why it's a necessity for searching internal documents.
  • Data pipeline stages: chunking strategies → vector embedding → storage → query retrieval → LLM generation.
  • Building a script with LangChain + ChromaDB (local vector database) + Gemini API.
  • Hands-on: ingesting private PDFs, querying contents locally, and printing outputs with citations.

Session 4 — End-to-End Research Workflow Lab

  • Integrated lab pipeline: NotebookLM (discovery) → RAG (deep query) → Overleaf (drafting) → AI assistants (polishing).
  • Participants work on compiling a brief, structured research note (1–2 pages).
  • Conducting peer-reviews in pairs to evaluate flow and citation accuracy.
  • Daily reflection (10 min quiz/feedback) & preview of Day 5 project guidelines.
Day 5 — Advanced Topics, Ethics & Capstone Theme: Multi-agent systems, responsible AI, deployment, capstone showcase. (Exploratory)

Session 1 — Multi-Agent Systems & the LLM Council Pattern

  • Why transition to multiple agents? Exploring task decomposition, specialization, and group debate.
  • Architectural typologies: hierarchical (supervisor control), collaborative (peer-to-peer), and competitive (debate).
  • LLM Council Pattern: Employing multiple LLMs to evaluate, critique, and vote on outputs to improve quality.
  • Hands-on demonstration: setting up a 3-agent council utilizing cloud Gemini, local Ollama (Llama 3.2), and local Ollama (Phi-3).

Session 2 — Responsible AI & Guardrails

  • Algorithmic bias, fairness, and toxic pattern mitigation in LLMs.
  • Hallucination: underlying causes, detection frameworks, and technical mitigation.
  • Live demonstration of security exploits: prompt injections and model jailbreaks.
  • Implementing guardrail scripts: validating inputs/outputs, content filters, and strict allow-lists.
  • LLM evaluation methodologies: assessing accuracy, groundedness, toxicity, latency, and operational cost.
  • Regulatory & privacy context: India's DPDP Act and an overview of the EU AI Act.
  • Academic ethics: guidelines on AI authorship disclosure, plagiarism checks, and peer-review policies.

Session 3 — Deployment on Free Platforms

  • Packaging Python applications for user distribution.
  • Streamlit Community Cloud: Deploying the interactive chatbot directly from a GitHub repository.
  • Hugging Face Spaces: Packaging and launching Gradio interfaces onto Hugging Face.
  • Best practices: developer secrets management, API rate limiting, and basic usage monitoring.
  • Capstone presentation polish (final 30 minutes in groups).

Session 4 — Capstone Showcase & Closing

  • Capstone showcase: Live mini-project presentations (3–5 min each, in teams of 2–3).
  • Facilitator and peer critique/feedback sessions.
  • Roadmap for continued learning: libraries, communities, and conferences.
  • Course feedback survey.
  • Certificate distribution ceremony and closing addresses.

Tools & SDKs Used (All Free)

Category Tool / SDK Free? Notes
LLM Chat (Web) ChatGPT, Claude, Gemini Yes Using free web tier interfaces
LLM API Google AI Studio (Gemini API) Yes Generous developer free tier key
LLM API Hugging Face Inference API Yes Access thousands of open-source models for free
Local LLM Ollama (Llama 3.2, Phi-3) Yes Runs model queries locally on consumer laptops
Orchestration LangChain & LangGraph Yes Open-source standard orchestration libraries
Multi-Agent CrewAI & AutoGen Yes Open-source framework for multi-agent workflows
ML Model Hub Hugging Face Hub Yes Hub hosting models, datasets, and demos
Interactive Notebooks Google Colab Yes Free web-hosted Jupyter environments with CPU/GPU
Vector Store ChromaDB Yes Open-source, local vector database
UI & App Prototyping Streamlit & Gradio Yes Fast, Python-only UI development libraries
App Deployment Streamlit Cloud & Hugging Face Spaces Yes Free online hosting for streamlit/gradio apps
Research & Writing NotebookLM, Overleaf Yes Free source research tool and online LaTeX editor
Academia Assistants Semantic Scholar, Connected Papers Yes Literature search engines and mapping tools
Visualizations Transformer Visualizer Yes Interactive attention visualizer at calm-sand-0f2fbf80f.7.azurestaticapps.net

Assessment & Certification

Daily Assessments

At the end of each day (Sessions 4), participants complete a 10-minute daily reflection quiz via Google Forms. It covers 5–7 multiple-choice or short-answer questions. The purpose is to reinforce learning and track gaps.

Capstone Mini-Project

Developed individually or in groups of 2–3. Options include:

  1. Domain Chatbot: RAG assistant for specific subjects.
  2. Research Agent: Multi-step paper survey agent.
  3. TA Agent: Automated quiz & rubric generator.
  4. Multi-Agent Workflow: Workflows using CrewAI/LangGraph.

Deliverables: Working GitHub repository with a detailed README and a 3-5 minute live demonstration on Day 5.

Certificate Eligibility

Certificates of completion will be issued to participants who meet the following requirements:

  • Maintain at least 90% attendance (4.5 of 5 days).
  • Submit at least 4 of the daily reflections.
  • Deliver and present their Capstone mini-project.

Logistics Checklist (for Organizers)

Infrastructure

Pre-Training Prep

Administration & Hospitality