This started with a phone call.
My buddy Jack Gorman floated an idea: what if you ran a bunch of AI models against a March Madness bracket and tracked how they did? Not as a gimmick -- as a real experiment. Different approaches, locked picks, verified results.
I couldn't stop thinking about it.
The Question
I work in AI consulting. I spend my days helping companies figure out what these tools can and can't do. But I'd never stress-tested them against something this clean -- 63 games, public data, binary outcomes, a universal scoring system. March Madness is a near-perfect AI benchmark hiding in plain sight.
So I wanted to know: do different AI approaches actually produce meaningfully different predictions? And if so, which ones hold up when the games start?
The Build
Standing up the site was the easy part. Claude Code and I got a working version live in under a week -- eight models, a scoring engine, bracket visualizations, a blog, a leaderboard, all running on Next.js and Vercel. The tool I used to build the thing is the thing I'm testing.
But "working" and "good" are different things.
The real time went into the grind after that. Iterations on iterations -- reformatting how brackets render, reworking the data structures until they were clean enough to trust, documenting every decision so I could retrace my steps. This wasn't "make me a cool March Madness site." This was build, look at it, realize something's off, adjust, plan the next piece, sometimes take a step back, then iterate again until it felt right.
Not perfect. I don't think it's perfect. But I got it to a place where I look at it and think: yeah, that's polished. That's something I'd actually share.
What I'm Trying to Learn
A few things, honestly:
How far can I push AI as a prediction tool? Not toy demos -- real predictions, locked in advance, scored against reality. No take-backs.
Where do models agree and disagree? When eight different approaches all pick the same team, that's a signal. When they diverge, the why is usually more interesting than the who.
Can this actually help someone win their pool? I built this hoping my friends can use it to fill out better brackets. Not mine, though -- I'll fill mine out the old-fashioned way and probably lose to someone who picked based on mascot vibes.
The Rules
Every model has an ESPN Tournament Challenge entry locked before tip-off. That's the timestamp proof. No retroactive edits. Results are sourced from ESPN -- never generated, never fabricated.
The models predict. The results are real. That's the whole integrity layer.
Thanks to Jack for the idea. Let's see what happens.