LLMs in the application layer
Since LLMs do not natively enjoy video (digesting every frame bloats context, most LLMs do not even accept audio input), I found the pursuit of composing LLMs in the application layer to improve output quality in a space known for slop to be interesting.
A quick look at how SaralClip reframes screenshare content.
Open Source
I think that walking into a popular repository as an unaffiliated contributor and seeing where you stand is something every passionate SWE should try.
With LLMs, it's easier than ever to generate a working PR but harder than ever to actually have that PR land in real repositories (~20k+ stars). It's even harder to land PRs that go beyond straightforward implementations or simple bug fixes and actually audit the engineering rigor of a codebase. My favorite contributions are the ones that remove code, audit engineering practices, and dictate how code should be LLM reviewed.
There are also some other UX improvements of mine I liked, and I really enjoyed talking to users who filed issues that pertained to areas of the codebase I knew very well.
For the full list see all my merged PRs.
I have a few standard contributions to the AI SDK. They are OK.
For a full list of PRs see all my merged PRs.