HNHacker News
TopNewBestAskShowJobs

kbyatnal

2,179 karma · joined March 23, 2016

building extend.ai
submissionscomments
kbyatnal··on Mistral OCR
We're approaching the point where OCR becomes "solved" — very exciting! Any legacy vendors providing pure OCR are going to get steamrolled by these VLMs.

However IMO, there's still a large gap for businesses in going from raw OCR outputs —> document processing deployed in prod for mission-critical use cases. LLMs and VLMs aren't magic, and anyone who goes in expecting 100% automation is in for a surprise.

You still need to build and label datasets, orchestrate pipelines (classify -> split -> extract), detect uncertainty and correct with human-in-the-loop, fine-tune, and a lot more. You can certainly get close to full automation over time, but it's going to take time and effort. But the future is on the horizon!

Disclaimer: I started a LLM doc processing company to help companies solve problems in this space (https://extend.app/)

kbyatnal··on Ask HN: Who is hiring? (March 2025)
Extend | Senior/Staff Software Engineer | NYC | Full-time | $175k-$250k | 0.50% - 1% Equity

Extend is building a LLM-native document processing platform (massive market w/ low NPS existing solutions) (https://extend.app/)

Apply here: https://jobs.ashbyhq.com/extend/9d4d8974-bd9b-432d-84ec-8268

--------

Why you should consider joining:

- We're looking for engineer #4, lots of equity, culture of high impact and ownership, in-person in NYC

- We've accelerated revenue 15x this past year to 7-figures in ARR

- Our customers include companies like Chime, Flatiron Health, Brex, Opendoor, Checkr, and many more

- We're supporting customer and revenue metrics with 1/2 the team size of other startups, so everyone joining at this stage will have outsized impact

- Backed by YC, Homebrew, investors from OpenAI, and more

kbyatnal··on Ingesting PDFs and why Gemini 2.0 changes everything
We work with fortune 500s in sensitive industries (healthcare, fintech, etc). Our policies are:

- data is never shared between customers

- data never gets used for training

- we also configure data retention policies to auto-purge after a time period

kbyatnal··on Ingesting PDFs and why Gemini 2.0 changes everything
This is spot on, any legacy vendor focusing on a specific type of PDF is going to get obliterated by LLMs. The problem with using an off-the-shelf provider like this is, you get stuck with their data schema. With an LLM, you have full control over the schema meaning you can parse and extract much more unique data.

The problem then shifts from "can we extract this data from the PDF" to "how do we teach an LLM to extract the data we need, validate its performance, and deploy it with confidence into prod?"

You could improve your accuracy further by adding some chain-of-thought to your prompt btw. e.g. Make each field in your json schema have a `reasoning` field beforehand so the model can CoT how it got to its answer. If you want to take it to the next level, `citations` in our experience also improves performance (and when combined with bounding boxes, is powerful for human-in-the-loop tooling).

Disclaimer: I started an LLM doc processing infra company (https://extend.app/)

kbyatnal··on Ingesting PDFs and why Gemini 2.0 changes everything
It's clear that OCR & document parsing are going to be swallowed up by these multimodal models. The best representation of a document at the end of the day is an image.

I founded a doc processing company [1] and in our experience, a lot of the difficulty w/ deploying document processing into production is when accuracy requirements are high (> 97%). This is because OCR and parsing is only one part of the problem, and real world use cases need to bridge the gap between raw outputs and production-ready data.

This requires things like:

- state-of-the-art parsing powered by VLMs and OCR

- multi-step extraction powered by semantic chunking, bounding boxes, and citations

- processing modes for document parsing, classification, extraction, and splitting (e.g. long documents, or multi-document packages)

- tooling that lets nontechnical members quickly iterate, review results, and improve accuracy

- evaluation and benchmarking tools

- fine-tuning pipelines that turn reviewed corrections —> custom models

Very excited to get test and benchmark Gemini 2.0 in our product, very excited about the progress here.

[1] https://extend.app/

kbyatnal··on Ask HN: Who is hiring? (February 2025)
Extend | Senior/Staff Software Engineer | NYC | Full-time | | $175k-$250k | 0.50% - 1% Equity

Extend is building a next generation document processing platform (massive market w/ low NPS existing solutions) (https://extend.app/)

Apply here: https://jobs.ashbyhq.com/extend/9d4d8974-bd9b-432d-84ec-8268...

--------

Why you should consider joining:

- We're looking for engineer #3, lots of equity, culture of high impact and ownership, in-person in NYC

- We've accelerated revenue 15x this past year to 7-figures in ARR

- Our customers include companies like Chime, Flatiron Health, Brex, Opendoor, Checkr, and many more

- We're supporting customer and revenue metrics with 1/2 the team size of other startups, so everyone joining at this stage will have outsized impact

- Backed by YC, Homebrew, investors from OpenAI, and more

kbyatnal··on Show HN: LLM-aided OCR – Correcting Tesseract OCR errors with LLMs
"real improvements came from adjusting the prompts to make things clearer for the model, and not asking the model to do too much in a single pass"

This is spot on, and it's the same as how humans behave. If you give a human too many instructions at once, they won't follow all of them accurately.

I spend a lot of time thinking about LLMs + documents, and in my opinion, as the models get better, OCR is soon going to be a fully solved problem. The challenge then becomes explaining the ambiguity and intricacies of complex documents to AI models in an effective way, less so about the OCR capabilities itself.

disclaimer: I run a LLM document processing company called Extend (https://www.extend.app/).

kbyatnal··on Ask HN: What have you created that deserves a second chance on HN?
that among others! a lot of manual work + some automated scraping (there's about ~3000 forums in the index today)
kbyatnal··on Ask HN: What have you created that deserves a second chance on HN?
thank you! results ranking is very primitive today, but a lot of users have commented that they would prefer ordering by date (latest first) so I've been looking into that
kbyatnal··on Ask HN: What have you created that deserves a second chance on HN?
https://crowdview.ai - search engine for forums and discussion sites

Like many of you, I find Google results to be full of SEO spam and have resorted to adding "site:reddit.com" or "site:news.ycombinator.com" to all my queries (since 2015!). Otherwise, it's really hard to figure out "what does a genuine, real life human think about this thing?".

But limiting my results to just Reddit isn't ideal because so much great content exists elsewhere. Lots of great information and conversations have moved elsewhere, and niche forums are still alive on the web! But it's impossible to find these places because they rank so poorly on Google. So I built a search engine across a curated list of these, making sure to remove any kind of SEO junk (blog spam, listicles, etc).

There's also a chrome extension that surfaces these results alongside Google, so you don't have to remember to keep coming back.

Please try it out and share any feedback! (and if you're interested in this topic, join the Slack)

kbyatnal··on Ask HN: What's in for ChatGPT, Stable Diffusion, etc. after dust settles?
I saw something along these lines on Twitter and think it's very true:

the right way to think about ChatGPT is not "what can I do if I had API access to an intelligent person?" which is how most are treating it (because it's not that smart), but rather "what can I do with API access to 1000 people of medium intelligence?"

kbyatnal··on Metaphor Systems: A search engine based on generative AI
This is interesting! I wonder how different the results are from just indexing the contents of the page and semantically searching them (vs. trying to predict the next link). Have you tried anything like that?
kbyatnal··on The Age of PageRank Is Over
yup totally agree, but I think it's nearly impossible to compete with Google and make something much better for certain types of queries (e.g. factual info, location queries, etc). They have too much data and too much of a head start.

The best way to compete with them IMO is to choose a slice of queries that they perform very poorly at (like non-SEO spam results) and build something 10x better. But that brings up the challenge of getting users to remember to come back to your search engine.

It's tricky for sure, with no clear answer. But for the first time in a long time, I sense that we're approaching a tipping point for Google's dominance over the web.

kbyatnal··on The Age of PageRank Is Over
IMO this article misses the biggest issue with search engines today. It's less about any ranking algorithm (like PageRank), but rather the root cause is with indexing. No matter how much you fine tune your ranking system, if your index is filled with SEO junk and blogspam, you're fighting a losing battle from the start.

For some reason, a lot of these search engines like to brag about the number of documents in their index, which never made sense to me. Maybe it was true in the past, but on the modern web, larger index !== better results. In fact, I'd argue the opposite since you're much more likely to serve SEO spam.

That was my motivation to start hacking around on CrowdView (https://crew-rho.vercel.app), a search engine specifically for forums and discussion content (e.g. forums, discords, twitter, reddit, etc). It has a curated index (today, curated by yours truly) to remove SEO junk and help you figure out "what does a real, genuine human think about this think?"

kbyatnal··on Ask HN: HN-like forum for literature and cinema?
Not sure about the main site, but AskMetaFilter still has a high level of discussion content. I find myself checking it every so often.
kbyatnal··on Ask HN: HN-like forum for literature and cinema?
Finding new forums and communities like this is one of the reasons I built CrowdView: https://crew-rho.vercel.app/

It's a search engine specifically for forums and discussion content (think message boards, Discords, Twitter, and ofc Reddit). If you try some literature/cinema queries, I think you'll end up with something interesting.

I tried "favorite books all time" and found a few literature forums: booktalk.org, forums.onlinebookclub.org, bookcrossing.com/forum, kboards.com, worldliteratureforum.com, proz.com/forum/literature_poetry-22.html, forum.thegradcafe.com/forum/43-literature-and-rhetoric-and-composition/

kbyatnal··on Show HN: A search engine based on RSS feed
I was very excited by this and it's promising, but it needs a bit more work to be usable. I was hoping the RSS technique would help avoid the SEO and blogspam that Google is now filled with. Unfortunately, I still get quite a bit of that (my query was "espresso machine" and I got a bunch of listicles, etc).

I think applying a layer of curation on top would go a long way to fixing this.

kbyatnal··on Forum Channels: A space for organized conversations
If you post on a public Discord, I think you are accepting that your comments will be made public. I totally agree with you on private Discords though.
kbyatnal··on Forum Channels: A space for organized conversations
I totally agree - I find myself missing the Google discussions filter (if you remember what that was?) very often. It would let you search only across forums, message boards, and other places where actual discussions where taking place (as opposed to SEO and blog spam).

I actually built a small tool for myself to try and replicate this. It searches across a few discords, a curated index of forums and message boards, and a little bit of Twitter (while applying some advance parameters to filter out the marketing spam). Works pretty well!

If anyone else wants to give it a go (desktop only for now): https://crew-rho.vercel.app

(the forum search takes a little bit of time to load, but you'll see it if you scroll down)

kbyatnal··on Ask HN: Google Search down?
"500. That’s an error" on all search queries
kbyatnal··on Writing your own search engine is hard (2004)
any plans for an engineering blog? sounds like you and the team would have some interesting additions to the post!
kbyatnal··on How “let it fail” leads to simpler code
the best TLDR I've seen of this philosophy is - if you find yourself writing all over your codebase

try { ... } catch (e) { console.log(e) }

then you should probably just "let it fail" since you can't actually handle the error

kbyatnal··on How to build undo/redo in a multiplayer environment by Liveblocks
What a fantastic, well-written blog post - well done.

Out of curiosity, what's the max # of simultaneous connections per room that LiveBlocks can support? (it's hidden behind the enterprise signup flow today)

kbyatnal··on Arc Browser Company: Chrome and Safari face a new challenger
why are all these comments so negative? I, for one, am very excited that talented developers and designers are reimagining such a core part of everyone's daily workflow. The embedded calendar / music player (along with auto archived tabs) is very cool - not sure how well it'll work in practice, but it's clear the team has put their heart into this. And more competition for Google/Apple is always good.
kbyatnal··on Show HN: Vantage Autopilot – Save on AWS EC2 Costs
the value of what Vantage is building is not by solving a technical challenge, it's by solving a marketplace/network effects problem (something that is much harder to do)
kbyatnal··on Duck Chess
hah, what a fun variant. Some of the most fun I've ever had playing chess has been "bughouse" (aka tandem chess). It's basically 2v2 chess with the trick being, you and your partner play opposite colors. Whenever you capture a piece, you hand it to your partner, and they can drop it anywhere they like on their board.
kbyatnal··on Show HN: I built a service to help companies reduce AWS spend by 50%
Very interesting idea - I'd be wary of this rubbing the wrong way against AWS's TOS though.
kbyatnal··on Stripe Financial Connections
lots of respect for what you and the team have built at Plaid, but this is exactly the opportunity. As a developer who's worked at multiple fintechs and integrated with Plaid more times than I care to remember, it's an incredibly frustrating experience.

Ask any fintech and they'll tell you - Plaid is simultaneously the best and worst vendor they use. Best because there's no real alternative, but worst because it causes so, so, so many headaches with how unreliable the product is. The time spent building product workarounds at every company to account for Plaid issues is tremendous.

If Stripe thinks they can build something better, then I'd really love them to try.

Edit: William (co-founder of Plaid) seems to have deleted his comment, but it was basically accusing Stripe of repeatedly copying other companies.

kbyatnal··on Show HN: Multiplayer Demo Built with Elixir
If I understand this correctly, it's basically transmitting database changes over a websocket? If so, how does that work re: performance when it comes to something like live cursor positions?

Writing to the DB every 50ms for each cursor on the screen doesn't seem great for performance...or am I missing something?

kbyatnal··on Error Handling Is Hard
IMO Erlang/Elixir does it best with the "let it crash" mentality. Once you internalize and truly understand how it works, you won't want to do error handling in any other way. The tldr; handle errors you can handle, let everything else crash and rely on the runtime to bring you back online.

The zen of erlang (https://ferd.ca/the-zen-of-erlang.html) goes more in depth if you're curious.

← PreviousPage 2 of 5Next →