Ameya Joshi

Image gen | video gen | llm evals | vlm evals

[Ameya Joshi @ameyaajoshi]

An anecdote from when we were building Parse 2.0 - I thought forms and tables were structured documents that people should be able to just fill out. Turns out table boundaries are just suggestions, and forms come in 20 different shapes. Most interesting find was this set of old archived tax documents from the 1920s/30s! Find some examples in our benchmark - RealDoc-Bench.

[Kushal Byatnal @kushalbyatnal]

Over 1 billion PDFs are created every day, but your agents still can’t read them reliably.

Today we’re releasing Parse 2.0, the most accurate document parsing API in the world.

Extend already processes millions of pages daily for leading AI teams like Brex, Mercury, Opendoor, Flatiron Health, and hundreds of others. Now, its even better.

Parse 2.0 is SOTA quality on RealDoc-Bench, our open source benchmark that measures agent success rate on real world docs that agents actually encounter in production.

We trained Parse 2.0 on 1M+ pages of the hardest documents seen in production. Here’s how it stacks up:

- #1 in healthcare, real estate, logistics, and financial services
- 95.7% agent Q&A accuracy on 581 docs (next best: 92%)
- 0.847 F1 on layout (next best: 0.759)

Give it a try today and build production-ready document agents with Extend.

[Anjney Midha @AnjneyMidha]

fascinating to watch the Black Forest Labs team casually automate audi's industrial manufacturing operations with their video models

scaling laws for robotics are here

[Oleksii Kuchaiev @kuchaev]

Banning open-source AI would be a historic mistake — and a self-inflicted wound to U.S. AI leadership.
I've been in this field long enough to remember when GPT-2 was called "too dangerous to release." That didn't age well — and it drew heavy criticism from researchers even at the time. 1/5

[Ameya Joshi @ameyaajoshi]

I just read an amazing piece of technical content: https://t.co/BreHNJzjaO.

Tbh, the I am not really the best student of electrical systems, and it took me a really long time to build intuition around impedance and Smith Charts. This would have made it insanely easier. Great job @arenaphysica & Christopher Bryant!

[Ameya Joshi @ameyaajoshi]

@kushalbyatnal Very excited to see that the AI engineering crowdfets to experience the best mangoes of all time! I am having mega-FOMO

[Ameya Joshi @ameyaajoshi]

@kushal124 @coderabbitai @greptile I have been using @greptile through my work. Pretty happy with it on most days.

[Kushal Byatnal @kushalbyatnal]

today, we're open sourcing the Extend CLI

watch Fable use it to parse and extract an entire corpus of documents — end to end, without leaving the terminal

try it yourself: npm install -g @​extend-ai/cli

it comes with a built-in agent skill, so claude code/codex/cursor know how to drive it out of the box

[Kushal Byatnal @kushalbyatnal]

Introducing Extend UI — open-source components for document agents

- 14 components & examples for PDF, DOCX, and XLSX viewers, plus bounding box citations, file upload, e-signature, and more
- fully customizable
- MIT licensed

when we started, we tried every file viewer and document component library we could find

unfortunately, none of them had all the functionality (and polish) that we wanted, so we ended up building our own for @ExtendHQ

it was only ever meant to be internal, but enough customers kept asking for it that we decided to give it back to the community

it's useful for building agents, user-facing document flows, or internal tools

we use and maintain it for Extend ourselves, so it'll keep getting better over time (and it's battle tested on millions of pages running through our system every day)

it also works with design system agents like @magicpatterns for faster exploration and prototyping

available today on the @shadcn component registry! some examples in 🧵

[Ameya Joshi @ameyaajoshi]

@yashgth @kushalbyatnal Yes. A non-trivial section of our benchmark consists of handwriting examples. Our drill-down shows a 97.9% accuracy on QA.

[Ameya Joshi @ameyaajoshi]

@kyledcollins @kushalbyatnal That's a valid approach for some pdfs, but often pages are embedded as images. Additionally, we also need to run more processing for figures, tables, and form regions to preserve the visual elements and structure in documents. It's a very heavy tailed problem!

[Ameya Joshi @ameyaajoshi]

@AnishJoshi06 Point us to all of your documents!

[Ameya Joshi @ameyaajoshi]

@akhi20187 @kushalbyatnal @Eli Parse 2.0 outputs word-level confidence scores. In addition, our extraction api (which incidentally also leverages our new engine), provides field level scores. Would love for you to try both out!

[Ameya Joshi @ameyaajoshi]

@ChrisSpoke @kushalbyatnal yes! we do have a good set of architectural and engineering drawings in our training data. Here's an example:

[Ameya Joshi @ameyaajoshi]

@baseten Thanks guys, for always being there to support us. Best support ever

[Ameya Joshi @ameyaajoshi]

@the_doc_doctor @kushalbyatnal Table boundaries are just suggestions.

[Ameya Joshi @ameyaajoshi]

@kushalbyatnal You should try out overstory too if you get the chance! Really creative drinks!

[Ameya Joshi @ameyaajoshi]

I love simple experiments like this. Evals like this on the simpler end will help us build more scientifically rigorous models that will help in scientific discovery.

[Dimitris Papailiopoulos @DimitrisPapail]

https://t.co/V7BSUMCcQB

[Dimitris Papailiopoulos @DimitrisPapail]

https://t.co/V7BSUMCcQB