<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Hacker News: jojogh</title><link>https://news.ycombinator.com/user?id=jojogh</link><description>Hacker News RSS</description><docs>https://hnrss.org/</docs><generator>hnrss v2.1.1</generator><lastBuildDate>Thu, 19 Nov 2026 21:45:32 +0000</lastBuildDate><atom:link href="https://hnrss.org/user?id=jojogh" rel="self" type="application/rss+xml"></atom:link><item><title><![CDATA[New comment by jojogh in "[dead]"]]></title><description><![CDATA[
<p>NeoSpark today announced the integration of Nano Banana 2 — Google's latest Gemini 3.1 Flash Image model, offering users near-Pro level image generation quality while maintaining Flash-tier speed and pricing.<p>Key Highlights:<p>1-Second Generation: Faster than Nano Banana Pro, standard generation time under 2 seconds<p>4K Output: Native 2K generation with true 4K export support<p>Precision Text Rendering: ~90% text accuracy in AI images, supports marketing assets and UI design<p>Multi-Subject Consistency: Single workflow supports up to 5 characters + 14 distinct objects<p>Real-Time Search Grounding: Leverages Google Search to accurately render specific logos, landmarks, and products
Pricing (credit-based, pay-as-you-go):<p>Nano Banana 2: $0.080/image (official) / $0.080/image (useneospark.com channel)<p>New users receive 50 free credits upon signup
Credits never expire
NeoSpark Platform Features:<p>Supports multiple top-tier models: Midjourney v7, Nano Banana Pro, Z-Image Turbo, Seedance 2.0, Sora 2, Veo 3, and more<p>No design skills required, AI Agent automatically understands your needs<p>Editable vector output, not static images<p>Commercial-use safe, no copyright risks<p>Use Cases:<p>E-commerce product showcase images and advertising videos<p>Social media marketing content<p>Posters, flyers, infographics<p>Blog covers and UI/UX wireframes<p>Discussion: Has anyone tried Nano Banana 2 yet? Compared to the Pro version, is the quality gap really "almost invisible" as Google claims?</p>
]]></description><pubDate>Tue, 03 Mar 2026 06:38:26 +0000</pubDate><link>https://news.ycombinator.com/item?id=47228935</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=47228935</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=47228935</guid></item><item><title><![CDATA[New comment by jojogh in "Is Canva too manual and slow in the age of generative AI?"]]></title><description><![CDATA[
<p>As a busy founder, my time is my most valuable asset. My days are a constant juggle between architecting our SaaS platform, leading our go-to-market strategy, and my family life. I have to work 60+ hour a week, often until 11 PM, only to wake up at 6:30 AM to get my kids to school before my workday starts at 8:00 AM.<p>I'm also deeply involved in our marketing, which means writing technical articles and sharing them on LinkedIn and X. For every post, I need an eye-catching cover image with a compelling title and beautiful graphics. For this, Canva.com has been an absolute lifesaver. Its massive library of templates and its simple, visual editor are brilliant. I love that it empowers someone like me, with zero Photoshop experience, to create professional-looking designs.<p>But with the dawn of the AIGC era, I've started to hit a wall with Canva.<p>Two things have changed:<p>The pace has accelerated. In the age of AI, everyone is leveraging tools to become more efficient. As a founder, I'm expected to do more and produce more in the same amount of time. Every tool in my workflow needs an AI upgrade to keep up. But Canva’s AI features feel limited. I still find myself using external AI tools for image edits, then manually typing and arranging text. A single social media graphic can still take me 1-2 hours. That’s just too slow now.<p>The possibilities have exploded. Powerful generative tools like Midjourney and nano banana have shown us the magic of natural language. We can now generate and edit incredible images in any style, just by typing. It’s a huge shame that this power isn't integrated into a practical, template-based workflow like Canva's.<p>I searched everywhere for a tool that bridged this gap—one that combined the structure of templates with the speed of natural language AI. I found nothing. So, I decided to build it myself.<p>With this new tool, a user can create professional designs through simple conversation. Imagine picking a template and, instead of manually editing, just typing in a chat box: "Create a design for a 4-day training course on AIGC prompt engineering" or "Change the background of this image to blue." The AI would instantly generate a beautiful design, automatically placing the text and images into a perfect layout. No design experience needed—just your idea and your words.<p>This is the "why" behind what I'm building. I'm solving a problem I face every single day.<p>Does this resonate with you? If you're a busy professional who believes AI can supercharge your creative workflow, I’d love for you to join me.<p>Please visit <a href="https://useneospark.com" rel="nofollow">https://useneospark.com</a> to join our wishlist. You'll be the first to know when our MVP goes live.<p>I'll be online and can't wait to hear your feedback. Thank you</p>
]]></description><pubDate>Fri, 14 Nov 2025 14:44:14 +0000</pubDate><link>https://news.ycombinator.com/item?id=45927187</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45927187</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45927187</guid></item><item><title><![CDATA[Is Canva too manual and slow in the age of generative AI?]]></title><description><![CDATA[
<p>Article URL: <a href="https://useneospark.com/">https://useneospark.com/</a></p>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=45927186">https://news.ycombinator.com/item?id=45927186</a></p>
<p>Points: 1</p>
<p># Comments: 1</p>
]]></description><pubDate>Fri, 14 Nov 2025 14:44:14 +0000</pubDate><link>https://useneospark.com/</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45927186</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45927186</guid></item><item><title><![CDATA[New comment by jojogh in "Ask HN: What Are You Working On? (Nov 2025)"]]></title><description><![CDATA[
<p>I am working on building neospark - AI native canva<p>And we are now recruiting our first group of seed users. If you're excited to shape the future of design and be among the first to try neospark, please visit our homepage and join the wishlist through <a href="https://useneospark.com/" rel="nofollow">https://useneospark.com/</a><p>We'll send you an exclusive invitation the moment our MVP is live.</p>
]]></description><pubDate>Fri, 14 Nov 2025 14:37:01 +0000</pubDate><link>https://news.ycombinator.com/item?id=45927089</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45927089</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45927089</guid></item><item><title><![CDATA[Show HN: UndatasIO's document parser MCP server is online]]></title><description><![CDATA[
<p>Hey HN, Alex here, founder of undatas.io.<p>Since we launched our core document parsing API, we've had fantastic feedback from developers who love the precision and the "pay-on-accept" model. But as we watched people build more complex systems, we noticed a common pattern: everyone was writing the same boilerplate code to manage their workflows.<p>Developers were building their own logic to handle workspaces, manage batches of files within tasks, poll for results, and handle state. Our raw API is powerful for single transactions, but it left the orchestration part up to you.<p>To solve this, we've built and are now launching our MCP (Multi-Channel Platform) server.<p>Think of it as a stateful, command-based orchestration layer that sits on top of our core parsing API. Instead of you having to write code to manage IDs and track the status of multi-file jobs, the MCP server handles it for you through a simple, logical interface.<p>The workflow is structured around a clear hierarchy: Workspace -> Task -> File.<p>You can use straightforward commands like:
- UnDatasIO_get_workspaces to get your work_id.
- UnDatasIO_upload to add files to a specific task_id.
- UnDatasIO_parse to kick off parsing for a list of file_ids.
- UnDatasIO_get_parse_result to poll for the job status without building your own loop.<p>This is perfect for anyone building complex, multi-file data processing pipelines, integrating with low-code platforms, or who simply wants to manage large batch jobs without the hassle of writing and maintaining state-management code. Our goal is to let you focus on what to do with the data, not the plumbing required to get it.<p>We believe this will make building robust document workflows much faster and more reliable. We're just getting started and are eager for your feedback.
We've put together a comprehensive guide and a video tutorial to walk you through it.<p>MCP User Guide (Full API Reference): <a href="https://docs.undatas.io/mcp/undatas-mcp/" rel="nofollow">https://docs.undatas.io/mcp/undatas-mcp/</a><p>Video Tutorial on YouTube: <a href="https://www.youtube.com/watch?v=_xobbKiPyf4" rel="nofollow">https://www.youtube.com/watch?v=_xobbKiPyf4</a><p>Our Official Website: <a href="https://undatas.io/" rel="nofollow">https://undatas.io/</a><p>I'll be here all day to answer any questions. Let us know what you think</p>
<hr>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=45741621">https://news.ycombinator.com/item?id=45741621</a></p>
<p>Points: 4</p>
<p># Comments: 0</p>
]]></description><pubDate>Wed, 29 Oct 2025 01:39:20 +0000</pubDate><link>https://docs.undatas.io/mcp/undatas-mcp/</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45741621</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45741621</guid></item><item><title><![CDATA[New comment by jojogh in "Parse complex documents in LangChain with new provider UndatasIO"]]></title><description><![CDATA[
<p>Hey HN, Alex here, founder of undatas.io.<p>Huge news: We just launched as a LangChain Core Provider—and we’re here to kill the #1 pain point of RAG: garbage document parsing.<p>Let’s cut to it: Building reliable AI used to feel like rolling the dice. Existing loaders mangle tables, drop critical data, and give zero way to verify outputs. You’d blindy feed messy text into embeddings, waste compute on garbage, and wonder why your app failed. I started Undatasio because this frustration broke more of my projects than I can count.<p>Our fix? Two non-negotiables: absolute parsing precision and total transparency—wrapped in a model no one else offers: pay only for the parses you accept. Bad output? It’s free. No excuses, no gotchas.<p>This isn’t "another loader" for LangChain. As a Core Provider, `UndatasioLoader` bakes quality control  into the start of your chain:  
- Programmatically check parsed JSON before it hits embeddings  
- Reject docs that miss key fields (e.g., no `invoice_total`, wrong table columns)  
- See exactly where data came from with positional `bbox` coordinates (build your own validation UI in minutes)<p>No more wasting time or money on downstream garbage. Data prep should be the reliable part of your stack—not the scary one.<p>We’ve been grinding to make this integration feel native to LangChain, and partnering with their team to push it live has been a blast.<p>If you’re tired of RAG failing because your inputs are broken, give it a spin. We’re here all day to answer questions, and we need your feedback to make this even better.<p>Links to get started are in the comments—fire away!<p>Here’s how to get started:<p>1. Install the Package: pip install langchain-undatasio
(PyPI Link: <a href="https://pypi.org/project/langchain-undatasio/" rel="nofollow">https://pypi.org/project/langchain-undatasio/</a>)<p>2. Check out the Official Docs:
(LangChain Provider Page: <a href="https://docs.langchain.com/oss/python/integrations/providers/undatasio" rel="nofollow">https://docs.langchain.com/oss/python/integrations/providers...</a>)<p>3. Try the Live Demo: We've set up a Colab notebook with examples.
(Google Colab Notebook: <a href="https://colab.research.google.com/drive/1k_UhPjNoiUXC7mkMOEIt_TPxFFlZ0JKT?usp=sharing" rel="nofollow">https://colab.research.google.com/drive/1k_UhPjNoiUXC7mkMOEI...</a>)<p>I'll be here all day to answer any questions. Let me know what you think.</p>
]]></description><pubDate>Wed, 24 Sep 2025 07:57:11 +0000</pubDate><link>https://news.ycombinator.com/item?id=45357539</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45357539</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45357539</guid></item><item><title><![CDATA[Parse complex documents in LangChain with new provider UndatasIO]]></title><description><![CDATA[
<p>Article URL: <a href="https://docs.langchain.com/oss/python/integrations/document_loaders/undatasio">https://docs.langchain.com/oss/python/integrations/document_loaders/undatasio</a></p>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=45357538">https://news.ycombinator.com/item?id=45357538</a></p>
<p>Points: 1</p>
<p># Comments: 1</p>
]]></description><pubDate>Wed, 24 Sep 2025 07:57:11 +0000</pubDate><link>https://docs.langchain.com/oss/python/integrations/document_loaders/undatasio</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45357538</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45357538</guid></item><item><title><![CDATA[Show HN: Undatas.io – A pay-on-accept document parsing API]]></title><description><![CDATA[
<p>Hey HN, Alex here, founder of undatas.io.<p>Our journey started from a place of deep frustration with RAG (Retrieval-Augmented Generation). I was helping companies build internal knowledge bases on their own data, and the promise was huge. But in practice, the results were often mediocre. Important information was frequently missed during retrieval, and we kept hitting dead ends.<p>After endless debugging, we realized the problem wasn't the LLM; it was classic "garbage in, garbage out." We traced the retrieval failures back to the very first step: document parsing.<p>Whether we used open-source libraries or expensive paid APIs, the story was the same. Precision was lost. Key phrases, critical numbers, and entire table rows would just vanish during the parsing process. We spent countless hours manually comparing the original PDFs to the parsed output to find what went wrong. It was a soul-crushing, time-consuming nightmare.<p>The biggest pain points were:<p>1. Complex Tables: Most tools collapsed when faced with real-world documents. Borderless tables, cells merged across rows and columns, or tables containing handwritten notes were consistently mangled.<p>2. Lack of a Feedback Loop: When the parser got something wrong, there was no easy way to manually annotate and correct it. You were stuck with the bad output.<p>I got so frustrated that I decided to build the tool I wished I had: a parsing engine obsessed with precision, that makes the entire data extraction process transparent.
That’s what undatas.io is. And today, we're launching our API.<p>We built our API around a simple principle: you only pay for results you actually accept.<p>To solve the transparency problem, every piece of extracted data in the JSON response includes its positional coordinates (bbox). This allows you to build your own "glass box" validator, mapping the data directly back to the source document, making the data prep stage for RAG completely transparent.<p>Our goal is to build the best and most trustworthy parsing tool for developers. We're just getting started and would be grateful for your feedback.<p>You can check out the docs and try it out here: <a href="https://doc.undatas.io/" rel="nofollow">https://doc.undatas.io/</a><p>I’ll be here all day to answer any questions. Let me know what you think.</p>
<hr>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=45111590">https://news.ycombinator.com/item?id=45111590</a></p>
<p>Points: 1</p>
<p># Comments: 0</p>
]]></description><pubDate>Wed, 03 Sep 2025 02:15:22 +0000</pubDate><link>https://undatas.io/</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=45111590</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=45111590</guid></item><item><title><![CDATA[New comment by jojogh in "Show HN: We built a document AI, you only pay for good data"]]></title><description><![CDATA[
<p>Thank you, John. Your professionalism and advice have been a great encouragement to us.</p>
]]></description><pubDate>Tue, 12 Aug 2025 09:40:47 +0000</pubDate><link>https://news.ycombinator.com/item?id=44874247</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=44874247</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=44874247</guid></item><item><title><![CDATA[Show HN: We built a document AI, you only pay for good data]]></title><description><![CDATA[
<p>Hi HN, I’m one of the creators of Undatas.io.
We've been working with document AI for a while and got tired of the standard model: upload sensitive files to a third-party server, send them to a black-box API, and pay for every page, even if the output is garbage. We decided to build a platform that fixes this from first principles.
Today we're launching V3, which is a major overhaul focused on data control, verifiable results, and a fair pricing model.
Here’s a breakdown of the technical approach:
1. Securely Process from Your Cloud Storage (S3, Box, Dropbox, GCS, Azure):
Instead of uploading files to us, you connect your own cloud storage. The architecture works via native integrations. For major providers like AWS, GCP, and Azure, you grant our system secure, temporary access using scoped-down credentials (e.g., an IAM role you control). For services like Box and Dropbox, you connect your account via a standard OAuth 2.0 flow, granting read-only permissions. In all cases, the principle is the same: our workers fetch the object for in-memory processing and it's immediately discarded. Your documents never land on our persistent storage.
2. "Glass Box" Visual Validation:
To solve the black-box problem, we built an interactive workspace. The backend returns JSON with detailed bounding box coordinates for every extracted token, line, and table cell. Our frontend uses these coordinates to map the structured data back to the source document image, allowing you to click on any JSON element and see it instantly highlighted.
You can see a live, no-signup demo of the UI here: <a href="https://undatas.io/" rel="nofollow">https://undatas.io/</a>
3. State-of-the-Art Table Extraction:
This was our biggest R&D effort. Most tools fail at complex tables (merged cells, nested headers, no borders). Our model moves beyond simple heuristics. It uses a hybrid approach, combining a vision transformer (ViT) to understand the visual layout with graph neural networks (GNNs) to reconstruct the logical cell-to-cell relationships. This allows it to correctly parse table structures that would otherwise be ambiguous.
4. "Pay for Quality" API:
This is built into our API workflow. When you process a document, the results for each page enter a "pending" state. You use the visual validator (or an approval webhook) to review them. Only when you explicitly "accept" a page's results is the transaction committed and your credits are used. A "discard" call costs nothing.
We're trying to be as open as possible. While the core engine is proprietary, we are open-sourcing our client SDKs and other tools.
Link to try the full platform: <a href="https://undatas.io/" rel="nofollow">https://undatas.io/</a>
A signup is needed for the full platform to manage API keys and credits. To make it easy for the HN community to test everything, we’re giving everyone 5,000 free credits (good for ~1,000 pages) and a 7-day trial of all features, including the private cloud connections.
As a thank you to early adopters from this community, the first 50 people who subscribe to a paid plan will get a 25% lifetime discount.
We’re here all day to answer your questions. Thanks for checking it out.</p>
<hr>
<p>Comments URL: <a href="https://news.ycombinator.com/item?id=44825752">https://news.ycombinator.com/item?id=44825752</a></p>
<p>Points: 1</p>
<p># Comments: 2</p>
]]></description><pubDate>Thu, 07 Aug 2025 15:28:00 +0000</pubDate><link>https://undatas.io/</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=44825752</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=44825752</guid></item><item><title><![CDATA[New comment by jojogh in "Mistral OCR"]]></title><description><![CDATA[
<p>High accuracy is the goal! But the multimodal approach introduces some complexities that can impact real-world performance. We break it down in our review: <a href="https://undatas.io/blog/posts/in-depth-review-of-mistral-ocr-a-pdf-parsing-powerhouse-tailored-for-the-ai-era/" rel="nofollow">https://undatas.io/blog/posts/in-depth-review-of-mistral-ocr...</a> 
As for use cases, it really depends on how well it handles edge cases…</p>
]]></description><pubDate>Thu, 13 Mar 2025 03:51:06 +0000</pubDate><link>https://news.ycombinator.com/item?id=43350220</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=43350220</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=43350220</guid></item><item><title><![CDATA[New comment by jojogh in "Show HN: AI that deeply understands you codebase"]]></title><description><![CDATA[
<p>what is the different with aider project, I thought they are doing the same work with you.</p>
]]></description><pubDate>Sun, 15 Dec 2024 02:56:54 +0000</pubDate><link>https://news.ycombinator.com/item?id=42421149</link><dc:creator>jojogh</dc:creator><comments>https://news.ycombinator.com/item?id=42421149</comments><guid isPermaLink="false">https://news.ycombinator.com/item?id=42421149</guid></item></channel></rss>