An Intro to AI Engineering

The Rise of a New Craft

I fine-tuned an LLM into a chatbot in a weekend for my Quantvesting framework — something that would’ve taken months just a few years ago. I didn’t write any training code. That’s when it hit me: AI Engineering isn’t just a new way to build. It’s a new way to think.

We’re entering a new era in software development — one where developers are no longer writing every line of logic, but instead collaborating with large, intelligent models that generate, predict, and interact. It’s not just a tech shift; it’s a mindset shift. Welcome to AI Engineering.

Why “AI Engineering” — Not Just “LLM Ops” or “GenAI Ops”?

You might hear terms like LLM Ops or GenAI Ops thrown around in technical circles. But these don’t quite capture what’s really happening. Operating a model is one thing. But building end-to-end intelligent applications — that’s a whole new game.

AI Engineers today are:

• Designing user interfaces for human–AI collaboration
• Crafting prompts instead of pipelines
• Building on top of powerful pre-trained models
• Thinking like product designers as much as programmers

This isn’t just model ops. It’s productized intelligence.

Foundation Models: The Brains We Don’t Fully Understand

At the heart of AI Engineering are foundation models — massive pre-trained models like GPT-4 or Mistral that can do everything from summarizing documents to writing code. They’re called “foundation” because they support a wide variety of applications without needing to be retrained from scratch.

These models come in two main flavours:

• Masked Language Models (MLMs) like BERT — great for understanding.
• Autoregressive Language Models (ARLMs) like GPT — great for generating.

But here’s the twist: Most developers using these models don’t fully understand how they work under the hood. And yet, they build with them every day.

That’s both powerful — and risky.

From Machine Learning to AI Engineering: The Mental Shift

Traditional ML Engineering starts with:
Data → Cleaning → Training → Evaluation → Deployment

AI Engineering flips that:
Product Idea → Prototype with Pretrained Model → Evaluate → Iterate

The shift isn’t just technical — it’s philosophical.

You don’t start with data. You start with intent.

You don’t chase accuracy. You chase utility.

And perhaps most importantly:
You think like a product person, not just a data scientist.

This is where craftsmanship returns — but it looks different now.

The AI Engineering Stack: A New Kind of Full-Stack

Let’s simplify it into three layers:

1. Application Layer

Where most AI Engineers live:

• Prompt design
• UX + interface engineering
• Evaluation of outputs

2. Model Development Layer

For researchers and experts:

• Fine-tuning models
• LoRA, Transformers, optimization

3. Infrastructure Layer

Managed by MLOps/DevOps:

• Model serving
• GPU management
• Cost/performance tuning

The stack is horizontal, but the magic is in how you orchestrate the layers.

The 80/20 Paradox: Why Finishing Is Harder Than Starting

Getting a working AI prototype today is fast.

That weekend chatbot I built for my investing framework? Done in hours.

But taking it from 80% to 100%?
That’s where the real work — and intelligence — kicks in.

Unlike traditional software, AI systems:

• Are non-deterministic (same input can yield different outputs)
• Have no absolute truth (answers are probabilistic, not binary)
• Are hard to test (unit tests don’t apply cleanly)

Engineering becomes less about control and more about designing within uncertainty.

The Risk of Shallow AI: Are We Losing the Craft?

There’s a quiet risk brewing in the AI boom:
The rise of shallow AI — building apps by stitching APIs without truly understanding the foundations.

It’s reminiscent of how WYSIWYG editors democratized web design, but also distanced many from the core craft.

Or how calculators liberated math, but weakened numeracy.

Today’s AI engineers must ask:

• Are we creating meaningful tools, or just prototypes that demo well?
• Are we thinking through edge cases, failure modes, and ethics?
• Do we understand the implications of building on black-box foundations?

As Steve Jobs said, “Real artists ship” — but real artists also care about the medium.

The Enterprise Reality: APIs Are Easy, Culture Is Hard

Many enterprises are rushing to integrate AI — but most are still in the API consumption phase. Few are redesigning their workflows, data systems, or team structures to embrace this new paradigm.

The real unlock lies not in calling GPT via an endpoint, but in:

• Empowering tiger teams that mix engineers, designers, and PMs
• Reimagining user experiences around intelligence, not features
• Building muscle for ambiguity, iteration, and prompt-centric design

As Clay Christensen taught us, true disruption doesn’t start at the core. It begins at the edge — with small, agile teams building new value.

Final Thoughts: A New Engineering Instinct

We don’t need every developer to become an ML expert.
But we do need them to grow new instincts:

• Product-first thinking
• Uncertainty-friendly design
• Curiosity about behavior, not just logic

AI Engineering isn’t about perfection. It’s about embracing emergence.

It’s a new craft — one that mixes code with context, prediction with design, and intelligence with intent.

And the best part?

You don’t need permission to start. You just need curiosity — and a weekend.

Have a good day,

— AR