Pulpie: Open-Source HTML Cleaning Models That Cost 20x Less Than Leading Tools

·1 views

Shreyash, founder of AI company Feyn, has released Pulpie, a family of open-source models designed to strip boilerplate elements like ads, footers, and sidebars from raw HTML pages. Unlike leading extraction tools that use decoder architectures generating output token by token, Pulpie uses an encoder approach that labels each HTML block in a single forward pass, making it compute-bound rather than memory-bound. This architectural difference allows Pulpie to clean one billion webpages for around $7,900, compared to roughly $159,000 with the current leading extractor, Dripper. The project was motivated by a real incident during development of a research tool, where an ad slipped through noisy web content and appeared in an AI-generated answer served to users. All Pulpie models are publicly available on Hugging Face, and a live side-by-side comparison with competing tools can be tested online.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

What Is an MCP Server? Anthropic's Open Standard Explained for AI Builders

Model Context Protocol (MCP) is an open standard introduced by Anthropic in late 2024 that allows AI models to connect with external tools and data sources in a standardized way. An MCP server is a lightweight program that exposes capabilities — such as querying databases or calling APIs — to any AI client that supports the protocol. Before MCP, developers had to write separate custom integrations for each AI client, meaning one capability could require six different implementations across six tools. MCP replaces this fragmented approach with a single protocol, so one server can serve all compatible clients including Claude, Cursor, VS Code, and custom agents. The standard supports tool discovery, tool invocation, resource exposure, and model sampling, making it particularly useful for teams building shared AI tooling at scale.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Companies Need Central Oversight of AI Agent Token Spend

Engineers across most organizations are independently building and tuning their own AI agent setups, creating fragmented token spending that no single dashboard tracks. The cost gap between a well-optimized agent loop and a poorly configured one compounds across teams, yet most finance departments cannot report what agent token spend cost last quarter or whether it is rising faster than productivity. Beyond the financial problem lies a knowledge problem: when one engineer discovers a prompt or skill that cuts token usage significantly, that insight stays siloed on their machine and never benefits the wider team. The article argues that centralizing AI agent infrastructure — including shared skills, prompt patterns, and model context protocols — allows improvements to propagate organization-wide and makes spend visible and manageable. Just as cloud cost discipline took years to mature, the author contends that companies must now build equivalent oversight muscles for agent token consumption before the invisible costs scale further.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Traliran AI Hub Aims to Unify Fragmented AI Workflows in the Browser

Developers working with multiple AI models often lose productivity to constant tab-switching, repeated authentication, and manual copy-pasting across fragmented provider interfaces. A new open-source project called Traliran AI Hub attempts to address this by offering a lightweight, serverless, client-side dashboard that runs entirely in the browser via GitHub Pages. The tool supports cloud providers such as OpenAI, Google Gemini, Claude, and Groq, as well as local model servers like Ollama and Llama.cpp, all accessible from a single interface. API keys and settings are stored locally in the browser, with no intermediary servers involved, which the project claims ensures user privacy. The hub also includes a multi-model querying feature, allowing users to send the same prompt to several models simultaneously for easier comparison.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BrowserAct Tops Product Hunt With 629 Votes, Sparking Real Talk on Browser Automation Limits

BrowserAct claimed the #1 Product of the Day spot on Product Hunt, earning 629 upvotes and finishing third for the week, outperforming tools from larger, established teams. The launch sparked significant developer discussion, with comments focusing on pain points like human handoff flows, DOM drift, CAPTCHA decay, and session persistence across IP changes. A cloud architect who had been running the tool in production for six weeks noted that BrowserAct's human-assist feature treats intervention as an intentional design choice rather than a failure state, allowing agents to resume seamlessly after manual steps like two-factor authentication. The tool's anti-detection system uses a three-layer approach combining browser fingerprinting, automatic CAPTCHA solving, and human fallback, which the architect reported held up consistently over the testing period at moderate request volumes. BrowserAct's GitHub repository also grew from roughly 2,300 to 3,600 stars around the time of the launch.

0 comments Read more at DEV Community