What Machines Actually See: A Glazing Contractor's First Look at Data and Features

You're sitting at your desk with a set of plans, a spec book, and a deadline. The architect wants a curtain wall bid by Friday. You flip through the drawings, scan the schedules, and start pulling numbers.

Here's what you don't realize: you're doing exactly what a machine learning system needs to do before it can "learn" anything.

You're converting chaos into structure.

This article is about that process. Not the fancy AI stuff—we'll get there. This is about the foundation: data and features. If you understand these two concepts, you'll understand 80% of what makes machine learning work.

And here's the good news: you already do this every day. You just call it something else.

When You Look at a Shop Drawing, What Do You Actually See?

Let's start with something concrete. You open a shop drawing for a storefront elevation. What's on the page?

Lines. Dimensions. Callouts. Notes. A title block. Maybe some section cuts. A glass schedule reference.

That's what's literally on the page.

But that's not what you see. You see:

•           Four openings, two fixed and two operable

•           1-inch IGUs throughout

•           A door with a closer and panic hardware

•           Aluminum framing that needs to hit a 4-inch sightline

•           A head condition that's going to be a pain to flash

You've done something remarkable: you converted raw visual information into structured, meaningful categories.

In machine learning, that raw visual information is called data. The structured, meaningful categories are called features.

Data = The raw stuff. Numbers, text, images, PDFs. Unorganized.

Features = The organized stuff. The specific attributes you extract from the data that actually matter for making decisions.

That's it. That's the foundation of machine learning. Everything else builds on this.

The Spec Sheet as a Dataset

Let's make this more concrete. Open any Division 08 spec. You'll find paragraphs like this:

08 44 13 - GLAZED ALUMINUM CURTAIN WALLS 2.3 INSULATING GLASS UNITS A. Provide insulating glass units composed of: 1. Outboard Lite: 1/4" heat-strengthened, Low-E coating 2. Airspace: 1/2" argon-filled 3. Inboard Lite: 1/4" tempered, clear B. Performance Requirements: 1. U-Factor: 0.28 maximum 2. SHGC: 0.25 maximum 3. VLT: 40% minimum

That paragraph is data. It's text. A machine can read the characters, but it has no idea what they mean.

Now, when you read that same paragraph, you mentally extract:

You just created a feature set. Each row is one feature. Each feature has a name (what you're measuring) and a value (the actual measurement).

This is exactly what machine learning needs. Not the paragraph. The table.

Two Types of Features (You Already Know This)

Look at that table again. Some values are numbers. Some are words.

•           1/4" — a number (you could do math with it)

•           Tempered — a category (you can't add "tempered" + "heat-strengthened")

In machine learning, these have names:

Continuous features = Numbers on a scale. Thickness, U-factor, square footage, price per SF.

Categorical features = Labels or categories. Coating type, manufacturer, frame finish, glass treatment.

You handle these differently in estimating, right?

Continuous: "The glass is 3/8" thick, and thicker glass costs more per SF. How much more?"

Categorical: "Is this Viracon or Guardian? Different vendors, different pricing tiers."

Machine learning handles them differently too. It needs to know which is which. If you tell a machine that "tempered" = 1 and "annealed" = 2, it might think annealed is "twice as much" as tempered. That's nonsense. Categories aren't quantities.

This is why humans still have to set up the data correctly. The machine can learn patterns, but only if you tell it what kind of feature it's looking at.

Why Raw Specs Aren't Ready for Machines

Here's a dirty secret: real construction data is a mess.

You already know this. You've seen:

•           Glass schedules that reference a detail that doesn't exist

•           Specs that contradict the drawings

•           Performance values on sheet A5.01 but makeup on G-101

•           Three different ways to write "1/4 inch" in the same document

•           Missing information everywhere

A machine can't handle this. If you give it raw spec PDFs and say "learn," it will fail.

Why? Because the data isn't structured. The features aren't extracted. The machine doesn't know that "1/4"", "0.25"", and "6mm" all mean the same thing.

Someone has to clean the data first. Someone has to decide what features matter. Someone has to organize the chaos.

That someone, for now, is usually a human. Eventually, AI will get better at this. But even then, it's doing the same job you do: turning messy documents into structured feature sets.

The Takeoff Sheet: You Already Build Feature Sets

Here's where it all comes together.

When you do a takeoff, what do you produce? A spreadsheet. Columns and rows. Each row is an opening or a system. Each column is a feature:

That's a dataset. Each row is a data point (one opening). Each column is a feature (one attribute of that opening).

If you wanted to train a machine to predict glass costs, this is exactly what you'd feed it. Not the plans. Not the specs. The takeoff sheet. The structured, organized, feature-rich version of the project.

"A machine learning model is only as good as the features you give it."

Garbage in, garbage out. If your features are wrong, incomplete, or poorly organized, the machine will learn the wrong patterns.

Sound familiar? It's the same reason a bad takeoff leads to a bad bid.

The Pipeline: Data → Features → Decisions

Here's the full picture, in both human and machine terms:

Human estimator:

Plans & specs (raw) → Mental extraction → Takeoff sheet (features) → Pricing logic → Bid

Machine learning:

Documents (raw) → Preprocessing → Feature set → Model → Prediction

Same pipeline. Different hardware.

The machine can process more data, faster. But it's doing the same fundamental job: extracting relevant features from raw information, then using those features to make decisions.

Once you see this, you can't unsee it. Every AI system you hear about—every "smart" tool, every "automated" solution—is just doing what you already do, at a different scale.

Summary: What You Need to Remember

Data is the raw material: specs, drawings, images, PDFs. Unstructured and messy.

Features are the structured attributes you extract: thickness, coating, U-factor, manufacturer. Organized and usable.

Continuous features are numbers (you can do math). Categorical features are labels (you can't).

Raw construction documents aren't ready for machines. Someone has to extract and organize the features first.

Your takeoff sheet is a feature set. You've been doing machine learning prep work your entire career.

In the next article, we'll explore feature representation—how machines convert those features into a format they can actually compute with. Spoiler: it involves turning everything into numbers, even the categories. And it's more intuitive than you think.

This article is part of a series on machine learning for glass and glazing professionals.