Joe Krebs

I’m seeing a lot of conversation around semantic graphs and context-aware AI. At Kurve, we’ve been building the foundation for this since day one. Our metadata graphs and automated multi-table data preparation layer are designed to give models the structured context they need to perform at their best. Combined with emerging tabular foundation models, this unlocks a powerful path forward for enterprise AI. The future isn’t just bigger models. It’s better structure. https://lnkd.in/guHMbWJe

14

Some cool data engineering techniques I have been doing to pre-filter the 1.3M files from the Epstein files to the correct processing pipelines. GLM-OCR is an incredibly accurate model - it has produced probably the most accurate text extraction of these files that exist - im looking forward to sharing these with the community, but this is only recognising text in the PDF files - it cant do much with the photographs, scans and forms that exist. My plan is to process these through a VLM to get an image description. This will mean it can take a photograph and return a description like: "This image captures a lavish, opulent interior, likely a grand mansion or a high-end private residence, characterized by its rich textures, dramatic lighting, and a striking, surreal element: a live tiger.  1. People Visible: There are no people visible in the image...  2. Setting and Location Context: A deep, velvety blue ceiling, possibly painted to resemble a night sky... Walls are lined  with dark wood paneling and intricate, gilded moldings...  3. Notable Objects: A large, live tiger is lying on its side on the rug, appearing relaxed and at ease. A large crystal  chandelier hangs from the ceiling...  4. Overall Mood: The mood is one of extreme luxury, grandeur, and a touch of surrealism or unease." But how do you do this? How do you go through a dataset that is primarily scanned PDFs, those PDFs can contain purely text content, full page images, full page colour scans, full page greyscale scans, scans of documents, forms etc. We could use a classifier model but for 1.3M PDFs this is going to take an age even with running it on a huge cluster. I figured that we could do this pretty low tech if we make a few assumptions and do a few passes: Pass 1 — Fast metadata scan (every page):  - Extract text length via page.get_text()  - Get image metadata: colorspace, dimensions, bpc  Pass 2 — Pixel analysis (only low-text pages with color-capable colorspaces):  - Skip pixel extraction entirely if text > 300 chars (definitely a text document extracted by the OCR process)  - Skip if colorspace is DeviceGray (literally cannot contain color)  - For Indexed/RGB/CMYK/ICCBased images on low-text pages: extract the image, thumbnail to 200x200, compute saturation (HSV) and channel difference (how much R/G/B diverge) Then put some basic thresholds in place like "if the saturation is above 5% and the amount of white (or almost white) is under 15%" then its likely a photograph. Blunt instrument for sure but it has taken down potentially 1.3M PDFs needing VLM treatment down to 86K files - much better. Will get these processing over the next few days as it will take much longer but this should be a fairly unique dataset that I've not seen out in the wild yet! #epstein #dataengineering #vlm #AI #ML #visualprocessing #GPU

4

Excited to see Databricks One becoming public preview. This is one of the top 5 features I've seen a need for across multiple organizations, and now, happy to see it become a reality. In addition to checking out Richard's post and article, check out the recording that I did with Ken Wong covering details that decision makers might find relevant: https://lnkd.in/e-wpUxVb

34

1 Comment