Pranay Agrawal

#100DaysOfCode My current project is to build an RL system which can play a card game called '28' at a superhuman level.

My video its officially out, been working on it for a longg time : youtu.be/x42JGBfRcXM

Continuously doing video editing now, will release the video sometime next week

Day 120: Forgot to update a few days earlier: Played the bot against my family and it absolutely crushed them, even with my local laptop without deploying on AWS/GCP for more CPUs. This is a sneak peek of the yt video im gonna make about my project:

Day 119: Found out about the AWS compute limits so switched to Google Cloud instead to get access to more no. of vCPUs.

Day 118: Deployed on AWS to scale the system to 500 vCPU to create an absolute monster of a model 😈.

Day 117: Fixed 2 out of the last 3 bugs related to card positioning and trick completion fading.

Day 116: Fixed the bugsrelated to card persistence and state updation. Last few will be done tomorrow.

Day 115: I'm back!! Started with the fix of the UI bugs today.

Exams starting tomorrow, will continue in first week of Feb

Day 114: The new strategy worked for the most part, will still have to iron out bugs though.

Day 113: New pure UI strategy failed too, trying one last time with an existing repository.

Day 112: Although UI was decent, Multiplayer interaction + lobby features failed miserably when using claude sonnet, switching to a different strategy tomorrow.

Day 111: Finally made good progress with the UI using claude 4.5 sonnet, will continue tomorrow.

Day 110(Forgot to upload yesterday): Started UI work, progress was slow. will continue tomorrow.

Day 109: Completely finished the bidding system integration, the bot works end to end now and is ready to dominate 😈. UI work starts tomorrow. Most likely done by Thursday.

Day 108: Almost finished integration of the bidding system. Will be fully integrated tomorrow, then UI work starts.

Day 107: Fixed the training stability issues and got 0.84 macro f1 score w/o hyperparameter finetuning. Also explored a rules based approach for the bidding. Will implement the bidding system for the bot tomorrow. Post that the final UI work is pending.

Day 106: Cleaned the collected data completely and started training. There were some training stability issues. I got an idea for an interesting fix by reducing the no. of output classes from 14-20 to only 14-16 because no one calls beyond 16 in a game.

Day 105: Completed the data collection, will train the model for bidding tomorrow.

Day 104: Made a website for data collection to train the bot for the bidding phase. Shared it with family. Will finish data collection tomorrow. I cant share it here for spam reasons 🥲.