Web @nutrientdocs, Speaker, Artist, Pianist. Developed https://nitter.cf/t.co/h1LBemNdYA, raaga, moose and more.
Bengaluru, India
Joined April 2014
- Tweets5.2K
- Following1.6K
- Followers4K
- Likes5.3K
Pinned Tweet
Introducing Reference. 🎉🎉
An online tool to create reference images of different human postures. 🕺
getreference.vercel.app/
🖌️🎨🖼️
I wanted a net worth app without endless manual entry, so I built it for myself 🔨
Sumora brings Indian & US stocks, MFs, gold, bonds, FDs, NPS and real estate into one view.
Automatic imports. Daily portfolio updates. Only real estate is manual.
Ritesh Kumar retweeted
We just launched a new website to better reflect what we've been focusing on for the last year. Go check it out. nutrient.io
Looks like Claude’s social media team has joined openai.
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
Ritesh Kumar retweeted
We put a new family of form-field detectors on @huggingface: form-field-v1. Three models that find Text, Choice (checkbox/radio), and Signature fields on a page image. The number that matters is what happens when the form is no longer blank.
The prior best open detector, FFDetr, scores .461 mAP50-95 on empty forms. On filled forms it drops to .170. Once someone writes in the box, it loses the box.
Our form-field-v1-medium goes the other direction: .529 empty, .783 filled, .764 handwritten. That's 4.6x better than FFDetr on filled forms, same shared pycocotools scorer.
That direction matters because in actual enterprise use cases most forms arrive filled not blank. Think insurance claims, applications, intake packets, etc. An agent that can only find fields on the blank template can't read the documents that actually show up.
The first tier I'd try is the open one: form-field-v1-nano. 0.9M parameters, 3.7 MB, Apache-2.0, ~90 pages/sec on a laptop CPU with plain ONNX Runtime and on filled forms it beats every baseline with 1/38th the parameters of FFDetr.
Benchmark (5,914 pages, 126,938 fields, CC-BY-4.0), scorer, and leaderboard are all public.
Run your own detector against it and send a PR:
huggingface.co/spaces/nutrie…
We built a focused model just for reading form fields, and it beats GPT and every other big model at finding and labeling them, the exact thing that trips up agents filling out forms 📄
huggingface.co/nutrientdocs/…
Here's another example where our specialized document models are leading frontier VLMs on document AI.
If you've ever tried to get an agent to fill a real form, you have likely experience this failure. It reads the page fine, then misses half the fields because it can't put a
tight box on them.
So our crack ML team @nutrientdocs built a Form Field Detection VLM and put it on a public benchmark.
On typed-field F1 at IoU 0.5 over 100 real form pages and 701 fields - our two models take #1 and #2 on a 16-model board. Nutrient Hybrid 0.477, Nutrient VLM 0.406. Best frontier VLM (GPT-5.6 Sol) 0.357, and the rest of the frontier trails from there.
The point isn't "we beat GPT." General VLMs are incredible. They're just not built for precise form-field localization like a specialized model is from a company with years of expertise in the document space. That's the whole bet.
Everything's open. The model card, dataset, demo, code. Run it yourself: huggingface.co/nutrientdocs/… and give it a like if you find it useful.
Ritesh Kumar retweeted
Everyone says Claude Cowork already does PDFs. It doesn't it, it just installs some open-source library to read extracted text yet stays blind to the actual document. So it can't show you where on the page
anything is.
We shipped the @nutrientdocs PDF Editor for Claude Cowork. You can use it now for free.
The full viewer renders inline in the chat panel, so you and Claude are looking at the same page. 16 tools: open, annotate, fill forms, redact, search. Real native PDF objects, not overlays: they travel with the file and save straight back to the original on disk. Drive it yourself, ask Claude, or both in one session.
Files never leave your machine. No signup, no license key, connector is MIT.
Runs in Cowork, inside @claudeai Desktop.
Ritesh Kumar retweeted
The accuracy-vs-speed trade in PDF parsing just collapsed.
we benchmark our document CLI against the open-source field on Opendataloader's Bench: 200 docs, hand-annotated ground truth, every parser at its latest release.
And our new --vision tier PDF-to-Markdown now tops every accuracy metric:
overall: 0.933 vs docling's 0.892
reading order: 0.959 vs 0.905
table structure: 0.938 vs 0.933
headings: 0.868 vs 0.829
and it does that at 0.35 s/page against docling's 0.55.
Full layout analysis, table reconstruction, formulas, handwriting. all local, nothing leaves your machine.
The part I love: 1.3.0 ran this corpus at 1.05 s/page. one release of profiling and GPU work later and 1.3.1 runs it at 0.35 with accuracy identical to four decimal places. I even had to re-verified this on a second machine before saying any of this out loud. Boy has the @nutrient_eng team been cooking!
If your PDFs are digital-born you may not need the --vision tier at all. The default engine parses at 0.004 s/page and is free up to 1,000 docs a month, no key, no signup.
Ritesh Kumar retweeted
We just got caught up in an attack that abuses Sentry's public event ingestion to slip prompt-injection payloads into our error reports. Fun. Luckily caught in time. Be really careful with pointing your agents at any open reporting channel. nutrient.io/blog/emerging-th…
Ritesh Kumar retweeted
Have you noticed Anthropic's shipped pdf-viewer plugin never actually shows the PDF?
It's fundamentally broken and has been for a while.
So we at Nutrient built one that works.
nutrient.io/claude-desktop
Ritesh Kumar retweeted
It's so crazy the noise around PDF parsers. Everyone creates their own benchmark now to claim performative SOTA. Anyone that has worked with PDFs knows how massive the minefield of edge cases are. That's literally why we have the business we have had for 12 years and our layout engine has been worked on for over 20 years.
@opendatalo51205 's benchmark is a really good bench of 200 different real world documents and it can't even cover all the edge cases of PDF. When you hear claims of most accurate from companies that have only been at the PDF game for a couple years I would check it against a common benchmark. We do it all the time and there's a reason we can make the claims we make without creating our own bench that we can overfit to.
Ritesh Kumar retweeted
My brain dump for @CIOonline on what AI is doing to the software industry and how to cope. cio.com/article/4171890/ai-i…
Four key insights from the article 🧵👇
We’ve often wondered how easily an LLM can use our docs and SDK, so we built this tool to help answer that question. It’s open source and may be helpful for you as well.
github.com/PSPDFKit-labs/age…
Ritesh Kumar retweeted
Different VLMs have wildly different spatial abilities.
We benchmarked 5 models on identifying precise pixel coordinates originating from PDFs (23 polygon vertices).
GPT-5.4-mini is 3-4x more accurate than Claude Opus 4 at the same task. Extended thinking made Opus worse.
Ritesh Kumar retweeted
Fresh out of the Nutrient bakery - the fastest PDF to Makrdown converter around. 0.007s per page. 90x faster than docling, 37x faster than pymupdf4llm. Without any compromise on accuracy. github.com/PSPDFKit/pdf-to-m…
Fastest PDF to Markdown converter is here 📈
Shipping Day! We’re all proud and excited @nutrientdocs to officially release our PDF to Markdown converter. On our 200 document hand annotated markdown based benchmark it beats out every other major parser for speed and reading order while matching accuracy.
We are shipping it as a closed source, free to use for up to 1,000 documents per month, no license key required, and your documents stay completely local.
We’ve had early beta testers give us phenomenal feedback but please reach out and share publicly your experience with it.
Our future roadmap:
- adding our OCR Vision capabilities to it.
- supporting Windows x64
- speed increases (yeah, make it even faster)
- support for DOCX to md
- more based on feedback.
If you find it useful, please star the repo and tell your friends!
Ritesh Kumar retweeted
New feature in @nudocsai! Just tell the genie what was in your document to get an old version back. 🤯 Like "git log -S", but for WYSIWYG documents.
Ritesh Kumar retweeted
We've been cooking @nutrientdocs on a PDF, DOCX, and other common document file types converter to Markdown packaged as a simple CLI library. On all our benchmarks we've done so far it's the fastest available by a long shot, most token efficient, and on par with accuracy of the top libraries.
- 2x faster than liteparse that was just released.
- 5x faster than pymupdf
- 4x faster than markit who @badlogicgames just praised for its speed.
- 3x faster than Microsoft's markitdown.
We'll be releasing the converter as freemium closed sourced here soon. Stay tuned.
Let me know if your interested in early access.
