I built an AI system to analyze poker videos.
I'm in Vegas right now for the World Series of Poker to play the Ladies' No-Limit Hold'em Event tomorrow. For anyone not familiar with the WSOP — before you get too impressed by the name "World Series" — I didn't qualify for anything. You just pay a buy-in fee and get a spot in the tournament.
It feels great to be here, and I’m really excited to play.
But admittedly, I feel out of poker shape since things have been so busy work-wise.
So the past couple of weeks, I've been trying to cram in some last-minute training.
And because I couldn’t help myself, I ended up building an AI poker analyzer this weekend to study more efficiently…
Let’s dive in.
The Setup
I’ve been burning through the Hungry Horse Poker Videos by Marc Goone. Thank you, Eric Deng, for introducing me to the channel.
He does an amazing job analyzing hands and talking through strategy. And he's a creative player — I think when you (as a viewer) are at a beginner/intermediate stage, there's such a stretch to the next level, where you’re trying to find interesting spots to play that aren't always obvious.
But I feel like the question in the back of my mind whenever I’m watching the YouTube channel is: Am I learning? Is this information actually sinking in?
I've also been playing around with poker trainers and ChatGPT.
I'll train on Octopipoker, and if there's a hand where I'm wondering why they’re telling me I should raise full pot rather than a half pot raise, I'll take a screenshot and ask ChatGPT to break it down. It does a decent job.
So I started to wonder if I could further automate analysis for Marc’s videos since the hands he’s playing are real spots he’s thinking through.
And turns out — I was actually able to get something up and running!!
Building a hand analyzer
Marc did a challenge last year where he tried to average $100 an hour in $5/$5 low stakes. He’d stream himself playing. Here is an example video: $100/hr at LOW STAKES? Brad Owen Weighs In! | Ep 3
To analyze his hands, I created a system that goes through the video, looks for every new hand, and processes a report and analysis for each hand based on the screenshots and Marc’s commentary.
Here is how it works step by step —
Step 1: Get the transcript using the YouTube API.
Step 2: Extract screenshots of key moments in the hand. I used yt-dlp to get the video stream and ffmpeg-python to capture screenshots. For example, in the hand below, it grabbed six screenshots showing the action progression.
Step 3: Analyze the screenshots. Using OpenAI's vision model, I created a prompt to extract the relevant poker information from each image — cards, pot size, positions, betting action.
Step 4: Combine everything. The system merges the screenshot analysis with the transcript and adds commentary around Marc's decision-making process.
Another Example Hand
Here is another example of a hand where Marc went all-in on the river.
Final screenshot of the hand
Analysis
This is an interesting hand because Marc bets the turn to set up a potential river all-in if he hits his draw, but then gets check-raised by his opponent. And ends up hitting the flush.
The Results
It actually works pretty well!
I programmed this quickly because I was curious if it was even possible. Lots of room for improvement, but here's what I found —
What worked surprisingly well:
With prompt engineering, OpenAI understood poker fundamentals and knew what information to extract from screenshots.
It could follow betting sequences and identify action and decision points.
The transcript and visual analysis combination actually made sense most of the time.
What needs work:
It sometimes confused suits (kept saying "King of Hearts" instead of "King of Diamonds"), so I wouldn’t completely trust the output yet for every hand.
Combining screenshots with transcript timing isn't always perfect.
The analysis could be a bit better with additional prompt engineering.
Overall, it could use some more tweaking to be more robust, but I was pretty happy with where everything ended up.
What’s Next?
Tomorrow is the tournament!
I’ve pretty much run out of study time, but this was fun.
Wish me luck tomorrow. I have very low expectations, but hopefully I get lucky and the last-minute studying pays off 💰
Thanks for reading!
I’m a General Partner at Chapter One, an early-stage venture fund that invests $500K - $2M checks into pre-seed and seed-stage startups.
If you’re a founder building a company, please feel free to reach out on Twitter (@seidtweets) or Linkedin (https://www.linkedin.com/in/jamesin-seidel-5325b147/).