Learning the Language of AI: From Python to LLMs, Prompt Engineering & RAG

Stay curious. I believe curiosity is the beginning of every meaningful breakthrough. — Marvette Johnson

And then the emails started.

Within hours of enrolling in the program in early February, I received several “Welcome to the Program” messages, including instructions for accessing the learning portal, Olympus.

Okay. This was really happening.

And so, the journey began.

The pre-work and first steps

The program consisted of four courses, plus pre-work designed to build foundational knowledge and a common understanding of key concepts.

The learning blended asynchronous lessons and readings, weekly live mentored sessions with hands-on activities, and practical application through knowledge checks, quizzes, and graded projects. The expected commitment was about 8–10 hours per week.

In practice, that meant a combination of:

• Asynchronous learning: recorded video lectures with PowerPoint presentations and supplemental readings.

• Synchronous learning: weekly live mentored sessions featuring hands-on activities and master classes.

• Practical application: knowledge checks, graded quizzes, and graded projects that let us put what we were learning into practice.

You’ve probably noticed that I haven’t mentioned the Python pre-work. Well, that’s because…

Python: “This Chinese is Greek to me!”

In a scene from Popeye the Sailor Meets Ali Baba’s Forty Thieves (1937), Popeye tries to make sense of a menu written in an unfamiliar script (supposedly Arabic). Completely unable to decipher it, he famously remarks, “This Chinese is Greek to me.”

Popeye the Sailor Meets Ali Baba’s Forty Thieves (1937)

Popeye’s predicament pretty much sums up my reaction when I opened my first Jupyter notebook.

Between the syntax and semantics, the data types and variables, the operators and expressions, and the loops and conditionals… Oh my!

And my initial reaction?

Escape hatches and lifelines

Remember the escape hatch the program manager offered me during my consultation, my “get out of jail free” card? During those first weeks, I was seriously considering using it.

The Python pre-work had me so overwhelmed that I started thinking about people I knew who actually spoke this language fluently. I thought about several former colleagues from my SNL Financial days (now S&P Global): brilliant people with technical expertise I knew I could trust.

I was tempted to reach out and ask one of them to coach me, or perhaps even tutor me, through this phase of the program.

And, as it turns out, I did.

I reached out to Pradipta Kishore Chakrabarty, a former colleague who was kind enough to agree to meet with me.

There was just one small problem.

He had no idea why I wanted to meet with him. 😂

I hadn’t yet explained that I was knee-deep in Python, staring at code that might as well have been written in another language, and wondering whether I had made a terrible mistake by enrolling in an Agentic AI program.

I wasn’t looking for someone to do the work for me. I just needed someone who could help me make sense of what I was looking at.

Finally cracking the code

As tempted as I was to call it a day, and before I “phoned a friend,” something else happened. I reminded myself of my guiding principle: stay curious.

A big part of staying curious is approaching challenges with an open mind and finding creative solutions. With that in mind, I abandoned all previous plans, and set out to find a way to make Python understandable.

What finally helped me crack the code was thinking of coding as a conversation with a computer, and of Python as a spoken language. The syntax was the grammar. The semantics gave the language meaning. Once I stopped seeing the code as a wall of unfamiliar symbols and started treating it as a language I could learn, it became much less intimidating.

“Once I stopped seeing the code as a wall of unfamiliar symbols and started treating it as a language I could learn, it became much less intimidating.”

I was nowhere near fluent, but I had learned enough to survive and, more importantly, to keep going.

With the pre-work completed, it was time to officially start the program. Orientation was held on March 7, and on Saturday, March 14, my cohort began its weekly two-hour live sessions, from 5:30 to 7:30 PM.

From Python to LLMs, prompt engineering and RAG

With a functional understanding of Python in my toolkit, it was time to begin the first course: Prompt Engineering Foundations.

The course began with Python Refresher for Agentic AI, which focused on using an LLM (ChatGPT) to debug Python code, handle Python files (organizing, structuring, and reformatting them), and read from and write to files.

Next came Introduction to Large Language Models (LLMs), which took me beyond the broad AI landscape and into the mechanics of modern AI: how foundation models work, how transformers and embeddings support them, how models are trained and customized, and where their capabilities and vulnerabilities lie.

I began learning not just what AI could do, but why it could do it, and where it could get things wrong.

Basics of Prompt Engineering shifted my focus from understanding the model to understanding how to communicate with it effectively. I learned how the way I structure an instruction, provide context, define a role, break down a task, or constrain an output can influence the quality of the response.

I was introduced to LangChain, which began moving me beyond individual prompts and toward thinking about how multiple AI components could be connected and orchestrated. I was also introduced to Hugging Face, a platform for hosting, sharing, downloading, and testing models, and was given access to an OpenAI API key.

Then came Advanced Prompt Engineering Techniques & RAG, which added another important piece of the puzzle: retrieval. I learned how Retrieval-Augmented Generation combines an LLM with external information, from basic vector-based retrieval to more advanced and agentic approaches. Just as importantly, I began learning how to evaluate AI outputs for accuracy, relevance, completeness, groundedness, and other measures of quality.

The course ended with our first graded project.

Using RAG, financial data, organizational reports, and LLMs, the project brought these two perspectives together to create a more holistic view of an organization’s potential. Rather than looking only at what the numbers say today, or only at what an organization says about its future, it connected financial strength with innovation potential to support more informed, future-focused investment analysis.

For me, this was the first real opportunity to see the concepts I had been learning come together in a practical application.

Finishing strong

By the time the first course wrapped up in mid-April, I had been on this learning journey for more than eight weeks, and I had begun to see AI differently.

I no longer thought of it simply as a chatbot that generated answers to my questions. I was beginning to see the systems behind the interaction: models, prompts, data, retrieval, tools, and evaluation working together.

I was learning the language.

And I was beginning to understand that there was a lot more to the conversation than I had initially realized.

From generative AI to agentic AI

Up to this point, the program had focused primarily on generative AI: understanding large language models, learning to communicate with them effectively, and exploring how techniques like Retrieval-Augmented Generation could extend what they could do.

I was learning how AI could generate, retrieve, and respond.

But the next course was about to change the focus.

The program was moving from generative AI to agentic AI: from AI that responds to our instructions to AI systems designed to reason, plan, use tools, and take action toward a goal.

I didn’t fully understand what that meant yet.

But I was about to find out.

Up next in this series: The Moment AI Became Agentic.