Day 1- End To End Agentic AI Project With LLMOPS

The core takeaway is a practical, end-to-end blueprint for building an agentic AI application—specifically an AI-based trip planner—using LangGraph and then engineering it for real-world delivery with LLMOps practices. The trip planner is designed to coordinate multiple specialized AI agents (weather, attractions, hotel costs, currency conversion, itinerary planning, expense calculation, and final summaries) so the system can handle a complex workflow with real-time data. That matters because it turns “agent demos” into a structured project that can be extended, deployed, and maintained like production software.

The session frames agentic AI as a collaboration problem: multiple agents, each tied to different capabilities or LLM-backed functions, pass information along a workflow until a goal is reached. LangGraph is positioned as the mechanism to model that workflow as a graph—built from nodes and edges—where each node performs a step and passes outputs to the next. The discussion emphasizes the fundamental components needed to make this work: node definitions (what each step does, including prompt engineering), edge behavior (how execution flows), and state management via a state graph so the system can maintain context across steps.

Day 1 is focused on problem definition and project scaffolding. The chosen use case—an AI trip planner for any city worldwide using real-time data—breaks down into concrete tasks: fetch real-time weather, retrieve attractions/activities, estimate hotel costs across multiple days, convert currencies for worldwide travel, generate an itinerary, compute total expenses, and produce a user-facing summary. The workflow is meant to resemble the helpful chat experience users expect from travel sites like Goibibo or MakeMyTrip, but built as an agentic system.

To make the project runnable and maintainable, the session shifts into engineering fundamentals for LLMOps: using Python with modular coding, and managing dependencies and environments with UV (a package/environment manager described as a Rust-based wrapper around pip). The instructor walks through installing UV, initializing a project with `uv init`, and creating an isolated virtual environment with `uv venv` using a specific Python version. Commands like `uv python list` and `uv python install` are used to ensure the correct interpreter version is available before environment creation.

A major portion of Day 1 is spent on creating a clean repository structure that supports scaling. The proposed layout includes folders for agentic workflow logic, configuration (YAML), prompt library, tools (separate modules for currency conversion, place search, weather info, and calculations), utilities (model loading, config loading, converters, calculators), and application entry points (e.g., `app.py`). The session also covers local package installation via `setup.py`, adding `__init__.py` files for import hygiene, and keeping secrets in an `.env` file that is excluded from Git tracking.

Finally, the workflow is prepared for deployment and CI/CD in later days, with mention of GitHub Actions and the idea of integrating observability/experiment tracking (e.g., via LangSmith) by loading the right API keys. By the end of the session, participants are given a GitHub repository link to fork/clone, so subsequent classes can build on the same scaffold and implement the full agentic system step-by-step.