uri_merhav

Try DocuPipe. It blows Gemini out of the water in terms of accuracy in extracting . They also generate a page + bounding box for every extracted field.

Very cool demo.

But isn't the point of an app being able to maintain a sharable worldview and state?

This "improvised just for you" sounds like a natural fit for a text adventure for example, but I'm still grappling with how this might apply to a real life use case.

One can conceivably imagine imagine LLMs one day transforming states for sharing on-the-fly as well, but it does sound very fragile.

There's lots of hidden gotchas to this. Uploading a screenshot and asking an LLM to transcribe one page is generally ok. Give it a table that spans pages, or a 60 page doc, and you're in dire straits.

I cofounded DocuPanda to handle this issue specifically. Call me biased, but I do believe it's the best solution out there.

Gemini Flash 2.0 is impressive but it hardly captures all of the information in the PDF. It's great for getting vibes from the document or finding overall information in it. If you ask it to e.g. enumerate every line item from multiple tables in a long PDF it still falls flat (dropping some line items or entire sections etc). DocuPanda and to a lesser extent Unstrucutred handle this.

Very strong table extraction, and a developer-focused approach where the AI helps you figure out a good schema to systematically extract the same fields from documents in all layouts.

Awesome stuff, looks very cool. I liked their focus on reducing hallucinations, that stuff is often overlooked by people rushing into building AI proof of concepts.

Gets the job done. I think it's better than unscreen.com

Please try again. Slight outage due to sudden shift in traffic volume

I'm the site's founder and happy to take hear any and all feedback.

The underlying technology is called diffusion models. I'm sure many on HN have already hear of it - it's the technology underlying all these neat text to image results in DALLE-2, Imagen, and the open source counterpart called Stable Diffusion.

Getting this technology to embed real life images in arbitrary backgrounds has been quite the challenge, but it's already showing some pretty neat results which you can check out on the website. Currently we support Christmas setup, and white background with shadows and reflections. We'll add a plenty more background categories in the immediate future.

I'm the site's founder - happy to hear you're impressed with the background removal. I think it's pretty much on par with remove.bg, which are the most famous competitor in the category.

Couldn't help but notice you omitted any reference to the major technological achievement here in placing objects on arbitrary backgrounds. It failed to impress on your use case?

Kind of like remove.bg but with a friendlier pricepoint. It also does photo vectorization and super resolution.

Post on refit and all the Facebook groups. Unfortunately people’s patience with you runs out pretty much immediately doing this.

I’m also pushing people to give out referrals with limited success.

My SEO game is non existent. I tried a blog post or two with nothing to show for it. I should reach out to sales gurus and such that reach Etsy/amazon sellers how to sell. I’ve only made halfhearted efforts there.

I think that now that I have a pretty decent api going and some users for that, I should probably make a medium blog post about my journey. Might land some more api customers.

Previous ideas: 1. Automated proffers io always profile pic selector and optimizer. Like scam your Facebook with permissions and pick out the best photo for LinkedIn, potentially enchanting and cropping Ali g the way. Terrible idea because consumers generally never pay and getting people’s attention to this is going to be far more expensive than the value I can give you.

2. Automated tests for “broken” front end based on how the side renders on various devices. Pretty sure a machine learning model can recognize broken design. I still think it’s a good idea but potentially a pretty complicated sell!

3. The rest I’ll redact because I’m still toying with doing them.

I own Product Pix (https://www.proproductpix.org). It removes the background from product photos, with the intended audience being mostly people who sell stuff online and need to set their background against a white background.

It makes $1300/month right now, up from $0 6 months ago. Living in the Bay Area, that would put me well below the poverty line if it were my sole source of income, so I'm not gonna call it "successful" just yet.

How I got started: I do machine learning, and I methodically searched for places where people buy a service transactionally on platforms like Fiverr and that I think can be automated away (or greatly automated with human reviewers in the loop) with state of the art machine learning models. There are hundreds or thousands of such opportunities that individuals can solve on their own.

I'll be more comfortable giving sage advice once I've crossed the $10K/month threshold, but still I'd say a willingness to try a lot of shit out and get digging on stuff you have 0 familiarity with is mandatory. In this project I've had to learn javascript, frontend, photography, google ads campaign management etc.

Another tip I wish someone had told me is, build a pricing page from day one. The temptation to get _some_ signal you're useful to people will drive you to offer stuff for free, but that will end up getting you a lot of unwanted attention from people who will never ever pay.

Also here's a Hackernoon post about the making-of (I'm the author)

https://hackernoon.com/how-i-built-my-first-machine-learning...

fine, or don't!

I'm the author, AMA :-)