My motivation for creating CompterPoker.ai was feeling a bit overwhelmed by some of the professional poker tools out there for learning GTO play. For some tools, learning how to simply operate the tool itself felt like a second job. With ComputerPoker.ai players can play against bots themselves simulating GTO play to learn what it "feels like" to play GTO vs. GTO opponents without having to turn any knobs or dials (feedback is real-time as you play).
The Beta tester code for HN Users is: HackerNews2026. All feedback is welcome! Please send suggestions for improvement or bugs to contact@computerpoker.ai or alternatively leave a comment below. Any questions I will do my best to answer.
As for the product offering the website is designed to teach players how to play optimal poker strategy (GTO) in simulated Texas Hold 'Em poker tournaments. Our value proposition is that if you can consistently beat the bots then you will fare well in live poker tournaments (of course adjusting for your opponents' play).
In addition to GTO pre-flop quizzes and pre-flop charts, users have the ability to simulate poker tournaments from start-to-finish and get feedback on their decisions _in real-time_ in a fun and low-risk environment.
For those interested the tech stack is Django deployed on AWS via Terraform and SaltStack, the database uses a Postgres RDS backend, and the frontend uses HTMX with WebSockets via Django Channels and Redis (Nginx serving as reverse proxy with CloudFlare DNS and SSL). During the project I used Claude Code to aid with various boilerplate aspects of the code base including building out the repos for Terraform and SaltSack and of course speeding up Django development.
Users are graded pre-flop based on the covered pre-flop scenarios (two-ways only for now). Post-flop users are graded on a residual MLP PyTorch model. We have built an in-house solver in Rust using the discontented CFR++ algorithm. The PyTorch model approximates GTO play post-flop (again only two-ways currently) based on training data with raises, EV, and realistic ranges for OOP and IP players. Because the post-flop decisions are based on a model that will always be a work in progress I refer to these decisions as GTOA (or "GTO Approximate").
Version 8 of the PyTorch model is the first one that I am happy with and actually find it quite difficult to play against. If you manage to beat the bots please do let me know how many tries it took! For those curious the PyTorch params for the most recent run are below (I trained on a gaming PC via Linux WSL2 using an AMD GPU):
``` (venv_rocm_native) $ python3 pytorch_model.py --data-dir unified_v8_raises/ --aggression-embed --predict-ev --d-card 64 --d-model 256 --n-attn-layers 3 --n-trunk-layers 4 --dropout 0.2 --weight-decay 0.05 --epochs 30 --batch-size 4096 --lr 3e-4 --steps-per-epoch 10000 --output gto_model_v8.pt
Training on cuda … … … _30 epochs later_ Saved best model (val_loss=0.3865) Training complete! Best model saved to gto_model_v8.pt ```
The website is live in Beta mode as I gather feedback on how things are structured and work out any bugs/kinks. If you have any suggestions for improvements I’d love to hear them. Subscriptions are live so if anyone wanted to test the Stripe payment processing flow I certainly wouldn’t mind! ;-)
p.s. This is a side gig for me. I am currently looking for full-time work either fully remote or on-site based in London, UK (this LLC that runs ComputerPoker.ai operates out of USA but I am based full-time in the UK and authorized to work in both UK and USA). If you or someone you know is looking for a SRE with strong software engineering skills please let me know!
No comments yet.
