Hand Wave

The last few weeks have mostly been me fixing all the unglamorous stuff like code quality and performance.

I rebuilt a good chunk of the recognition pipeline, improved accuracy, fixed camera and streaming issues, added better logging, and spent way too much time getting CI and TestFlight releases working.

But, Hand Wave is finally on TestFlight now! Unfortunately Meta’s SDK prevents me from shipping to the App Store just yet, but this is as close to ‘shipped’ as I can get. Just gotta record a demo now.

Finished! Sat Codex on a 16hr goal loop to try and optimize the CTC beam, fixed the harness, got the mobile app looking great. Ready to ship :)

Finally got a working version (pictured here is the result after fingerspelling machine learning properly, with <100ms latency!).

Almost finished with the mobile app bridge that will let me connect the Meta AI glasses I have to the inference server I’m hosting on Modal, that’s running locally here. Got featured on NBC as well a few weeks ago – https://go.aadisanghvi.com/nbc

Soon will have the entire thing launched. The idea is to have a fully closed loop with just wearable smart glasses, which are, hopefully, gonna be huge tech in the near future. So in theory, I could look at someone that’s hearing impaired, watch them sign, and hear translations in my ear!

Currently limited by compute capacity on Modal (my inference provider), and my own machine learning model. Unfortunately wasn’t able to track the time I spent training the model, but I used the dataset from Google’s fingerspelling Kaggle competition. An early version is already shipped! Almost ready to record a launch vid.