jtap

I've also experimented with the Brazil purchase data from Olist. That was pretty fun because it has more than on csv. The tool managed pretty well and was able to come to statistically correct conclusions.

The goal hasn't changed too much, make building decks easier by knowing exactly what you own and where it’s stored. You organize cards into boxes, search your inventory, search friends’ collections, and keep track of trades instead of digging through a similar closet of cards that my daughter and I search for.

The fun part has been the AI. I trained computer vision models that run entirely on the phone to detect and identify Pokémon cards. Training has become the slowest part. For the model that needs to be retrained every new release, I’m up to about 5 hours per epoch on my M4 Mac with 16 GB of RAM.

The Android app is currently in public testing with people from my local Pokémon league. It’s built with React Native, and I’m working on the iPhone version next.

Still lots to build, mostly around product and ux, and because a recent stupid mistake on my part, backups and deployment safeguards.

It's been pretty fun to build. Trained a couple of models to identify a card that run on the phone app. They work very well and are comparable, and much better than many, to any other card identifier in the market. As the models get more complicated training is getting a bit longer. I'm at about 9 hours per epoch on my Mac M4 with 16G of memory.

The Android phone app is in public test with a couple of my local group testing. The app is built in React Native, and I’m hoping to get an iPhone version out soon since there are a bunch of iPhone peeps.

It's been a super fun tool to build. The phone app just got approved in the Android app store. I have a bit of cleanup, but plan on releasing it soon.

I’ve been building a phone app + website (https://MyBulkCards.com) to scan cards and organize where everything is. It’s pretty basic right now, but I can store cards in boxes like “Box 1 AAA, Box 1 BBB, …” and find cards easy peasy. There’s also a friends feature so I can see what others have locally. We borrow cards from each other quite a bit.

It’s been a fun project to build. I trained one model to find a card in the camera frame and another to identify it. Still iterating a lot. One epoch on my Mac M4 takes about 2 hours, and I’m still seeing improvements past epoch 10. Even now, it can find and identify a card more often than not, even without the OCR bits. Both models are under 20MB, run directly in the camera frame, and are fast enough to identify a card as I slide it into view.

I started with Android since that’s what I have, and I’ve shared the app store testing link with my local group for testing. The app is built in React Native, and I’m hoping to get an iPhone version out soon since there are a bunch of iPhone peeps. A couple of the players also got me into MTG, so now I’ve got a pile of Turtles cards too. I’ll be training an MTG model next. I don’t think it’ll be too bad since I can reuse most of the same approach.

I’ve been thinking, for a while now, how I could merge a LLM and an Expert System. What I wanted was the ease of use of an LLM and the rules based processing of an Expert System. Over the past couple of weeks, I built what I thought was a unique way of merging the two. I choose to use auto insurance as an example because of my past history in the industry, but I feel as though it would be applicable in other industries. It’s still very rough around the edges, but if you are interested you can try it out at https://mktbx.com/#tryout I’d appreciate any feedback and your thoughts.