Gen AI & Agentic AI Expert

How large language models work, and the Python you need to build with them.

• What generative AI is: tokens, embeddings, transformers and attention, explained without heavy maths
• How models are trained: pre-training, instruction tuning and RLHF, and why they hallucinate
• The model landscape: GPT, Claude, Gemini, Llama, Mistral and Indian models such as Sarvam and Krutrim
• Python for AI apps: virtual environments, requests, async basics, pandas and Jupyter
• Calling an LLM API: keys, tokens, context windows, temperature and cost per request
• Running open models locally with Ollama and Hugging Face
• Responsible use: privacy, copyright and the Indian DPDP Act for AI apps

Getting reliable, checkable answers from LLMs for real business tasks.

• Prompt patterns: roles, clear instructions, few-shot examples and step-by-step reasoning
• System prompts, guardrails and refusing out-of-scope questions
• Structured output: JSON schemas, function calling and validating answers with Pydantic
• Classification, extraction and summarisation at scale
• Working with images and PDFs: vision models and document understanding
• Testing prompts: test sets, regression checks and comparing models
• Reducing cost and latency: caching, batching and choosing the right model size

Answering questions from your own documents with sources, not guesses.

• Why RAG: grounding answers in documents and citing sources
• Chunking strategies, embeddings and similarity search
• Vector databases: FAISS, Chroma, pgvector and Qdrant
• Hybrid search, metadata filters and re-ranking
• Building RAG pipelines with LangChain and LlamaIndex
• Evaluating RAG: faithfulness, answer relevance and retrieval recall (RAGAS)
• Handling tables, scanned PDFs and Indian-language documents

When to fine-tune, how to do it cheaply, and how to prove the model got better.

• Prompting vs RAG vs fine-tuning: choosing the right approach
• Preparing instruction datasets and synthetic data
• Parameter-efficient fine-tuning with LoRA and QLoRA on a single GPU
• Small language models: Phi, Gemma and Llama variants for edge and low-cost use
• Quantisation and serving with vLLM and llama.cpp
• Evaluation: benchmarks, LLM-as-judge, human review and bias checks
• Tracking experiments with Weights & Biases or MLflow

Building AI agents that plan, use tools and work together to finish tasks.

• What makes an agent: the reason-act loop, tools, memory and planning
• Tool use and function calling: web search, databases, APIs and code execution
• Agent frameworks: LangGraph, CrewAI, AutoGen and the OpenAI and Claude agent SDKs
• Model Context Protocol (MCP): connecting agents to apps and data safely
• Multi-agent patterns: supervisor, hand-offs and debate
• Human-in-the-loop approvals, guardrails and stopping runaway agents
• Evaluating and debugging agents with traces (LangSmith, Langfuse)

Shipping Gen AI apps that are safe, observable and affordable, and your capstone project.

• Deploying LLM apps: FastAPI, Docker and cloud options (AWS Bedrock, Azure OpenAI, Vertex AI)
• LLMOps: prompt versioning, monitoring, cost dashboards and alerts
• Security: prompt injection, data leaks, jailbreaks and the OWASP Top 10 for LLMs
• Responsible AI: fairness, transparency and India's AI governance guidelines
• Designing a Gen AI product: user research, success metrics and pricing
• Capstone: plan, build, test and demo an end-to-end Gen AI or agentic app
• Career preparation: portfolio, GitHub, interviews and Gen AI job roles