Ideas
Libraries we might build next
Three open-source libraries we’re thinking about. Tell us which you’d use, or suggest one of your own.
The ideas
- The problem
- One web page can be megabytes of HTML, and most of it is noise to a model. Converting to Markdown saves space but drops tables, forms, links in context and metadata.
- What it would do
- Clean a page down to semantic HTML that keeps what matters. Find elements from a plain query, like “the submit button” or “items and prices,” and get back CSS selectors or a reusable template. A small model labels the page and code does the rest, so nothing is invented.
- The problem
- No .NET OCR library is both local and fully managed. The usual choices wrap native binaries, are commercial, run only on Windows or send your images to the cloud. And for invoices and forms, reading the text is only the start.
- What it would do
- Read text from images in pure C#, with small models that run well on a CPU. Understand documents too: which number is the total, which lines are items, which date is due. Use an LLM vision model when you want one. Built to work with oneimage.
- The problem
- Good address parsing today means a heavy native library or fragile, US-only patterns. Finding names, phone numbers and addresses in web pages and free text has the same gap.
- What it would do
- Turn an address, phone number, email or website into one canonical form, and find contact details in pages and free text. Small models we train ourselves, running offline, with every field traced back to the input. The US and Canada first, then the UK, Europe, Mexico and Australia.
We’ll only email you if we start building the one you picked.
What should we build?
Up to 1,000 characters. Leave an email if you’d like a reply.