Go on our discord :) We can help u to find an issue and some implementation that could be usefull as demo or others x)

A quick note on the implementation details for those interested in compilers:

The hardest part wasn't the AD itself, but managing memory safety during the "growth" phase. Since NOMA compiles to native code (LLVM), I had to ensure that when a weight buffer gets realloc'd (moved in memory):

The gradient tape updates its pointers.

The optimizer state (Adam moments) is correctly mapped to the new indices.

The benchmark I linked shows the result: "Preserving" this state allows the model to continue converging immediately after resizing, whereas "Resetting" it causes a massive performance regression.

I'm specifically curious if anyone here has experience with handling SSA Phi-nodes during reverse-mode AD on the Control Flow Graph? That's my next big hurdle for supporting complex control flow.

Looks cool !

Hello HN, OP here.

I built this TUI (Terminal User Interface) because I wanted to watch market trends without leaving my command line environment.

It uses Yahoo Finance and CoinGecko APIs with a local caching system (JSON) to handle rate limits gracefully. It is written in Node.js and uses `blessed` for rendering.

Happy to answer any questions about the implementation!