B.BRANDON FRANK
HAN-CHWEEN HUANG

Nothing ventured.
Nothing gained.

English proverb. An earlier version appears in John Heywood’s 1546 collection.

Who is Brandon?

I build AI and data systems. If a model gives me an answer I didn’t expect, I want to see the document it read, the data it used, and whether I can get the same result on my own machine.

What would happen to a galaxy’s stars if you took away its dark matter? And what exactly moves inside an old G-Wagen door when you pull the handle? You can try both here.

SCROLL TO BEGIN ↓
01 / THE QUESTION

What’s AI
good for?

Can it extract the right information from a messy document? Can it make a workflow faster? Those are the kinds of questions I work on.

TRY AN EXAMPLE ↓

What are
we missing?

An extracted answer is only useful if it matches the source. I work on getting information out of documents and checking it against what was actually there.

The source matters as much as the model.

What happens
if you change it?

Take the halo out of the galaxy model and watch the orbits change. Or open the G-Wagen door latch and watch the fork close around the striker.

EXPLORE MY CURIOSITIES ↗

02 / HOW I GOT HERE

And then I wanted
to know how it worked.

My work has taken me from data pipelines and document processing to local models and automated media tools. Here’s how those pieces connect.

01 — THE FOUNDATIONS

How does all
this connect?

I've spent a lot of time getting information from one place to another and making sure it still makes sense when it gets there.

That's taken me through data systems, document parsing, and plenty of debugging. Eventually I wanted to see what else those pieces could do together.

FILESEVENTSDATABASES
✳TRANSFORM
USEFUL
INFORMATION
01 / FOLLOWING THE INFORMATION
02 — UNDER THE HOOD

Can I run this
on my own machine?

I like being able to run a model myself and mess with it. Then I run into memory limits, speed, and the question of how much model I actually need.

I've been working with local language and vision models, fine-tuning, and evaluation to figure that out.

?THE TASK
MEMORYSPEEDQUALITY
02 / RUNNING A MODEL LOCALLY
03 — BRINGING IT TOGETHER

Okay, does the
whole thing work?

I've worked on tools that turn content into videos: narration, visuals, captions, publishing, the whole sequence.

There's a lot to get right. Do the words match the pictures? Are the numbers right? What happens if one step breaks? I want the finished thing to be worth using.

AN IDEA, TAKING SHAPE.
STORY
VOICE
VISUAL
03 / VIDEO TIMELINE

Got something in mind?

Anything you can imagine,
we can build it together.

If there’s a task you keep putting off, I’d like to hear what makes it such a pain.

You bring

Your world.

Show me where the work slows down. A spreadsheet you fix by hand every Friday tells me more than a pitch deck.

I bring

A way to build it.

I can build a prototype with your data. Then we can try it on a real example and see what still takes too long.

We don’t need the whole plan to start.

Or start here

Your thought goes into an email draft. You decide when to send it.

LET’S COMPARE NOTES.

WHAT HAVE YOU BEEN THINKING ABOUT?

There’s so much more
to explore.↗