theseamusjames

If there's ever been a perfect app to vibecode your own replacement, it's postman. Claude can knock out a completely usable competitor in under an hour, I'm sure of it. A day of tweaking and you'll have everything you need.

Try lopsy.art. I've been at it for 6 months. Updating and building out features almost daily. It's a Rust/WASM engine too. Also open source. github.com/theseamusjames/lopsy.art.

I think there are about a dozen of these now.

I did this a few months ago and it's not simple: different sensors have different pixel arrangements, files can have different data structures, and the default tone mapping is provided by the app (so opening the same raw file in darktable vs preview will have different results). There are different algorithms for how you assemble the raw data as well, and the issue is that you can get something _passable_ that isn't _optimal_, and it's really hard for the LLM to recognize that something isn't right. Is it the pixel processing, or is it the tone mapping, or is that just how the image looked?

And on top of that, you need a bunch of different files from different sensors to test with. Canon, sony, apple, fuji, nikon, olympus, etc. -- if you're building a truly general application, that's a concern.

Or if you're like me and you shoot fuji, you just build a fuji raw parser and call it good. The "TAM of 1" world of software development...

Yeah, it's a really well defined problem space with quantifiable outcomes, so LLMs are very well suited to it.

For lopsy, the biggest issue I ran into was combinatorial bugs: certain sequences of tools and actions would cause a bug to occur that wasn't obvious from isolated tool testing. I solved it by having Claude generate complete projects, full documents with novel concepts chosen randomly, built in the browser with playwright. This continues to expose so much more than individual unit/tool tests, and I have an automated loop that addresses them. (And a bonus from that: I use the outputs to make tutorials to show what the tool can do).

And 500 levels of org tree, and grumpy customers that freak out if you change anything about their workflows, and decades of legacy code across dozens of micro-services, and backwards compatibility requirements with 20 year old formats...

Raw was a tough one when I tried it a few months ago -- I never thought about how individual color pixels on the sensor come together to form a color value for a point in an image file. Super interesting, and not at all standard between the different manufacturers. It's a hard problem for software to solve in a general way.

I've been doing the same thing on lopsy.art.

Great news: we're excellent at curing mice of diseases. All this time we've been trying to find drugs that have the same effect in humans as they do in mice, but until now no one has had the insight or bravery to ask the question, "What if we just made people more mouse?" Exciting stuff.

But they're all converging on capability. Do I care if it's a 72% or 74% on SWEBench? Practically, probably not. And if I'm not paying per token locally, then if it takes a tiny bit longer to get to the result, I don't care.

Can't wait for the new qwen/deepseek/kimi releases 2 weeks from now.

There's been a spike in cases of people eating death cap mushrooms in California this year, and I would bet almost anything that LLM adoption has something to do with it.

I liken it to the auto-industry in America during the 1950s. Back then, single income families could take long summer vacations to cabins they might even own. The jobs came with good benefits and company culture was important. It feels very similar.