X (formerly Twitter) ·
📢 Excited to finally show y'all what we've been building at Reworkd! The Internet, a planet-wide meta-cortex, is man's greatest invention. But it's a jungle—to query web data, you need an API & an engineer. We use multimodal codegen to turn the web into structured data 🧵⬇️
Our codegen agent takes: 1) a schema for the structured data to extract 2) the target website's URL On the website, the agent then explores, extracts multimodal context, writes code, evaluates it, and iterates until its generated code robustly extracts the desired data.
We've open-sourced much of our stack. One major finding is that stringified web UIs eventually outperform HTML and screenshots as context for codegen. We built Tarsier (1.2k+ stars) to provide this multimodal context to our agents & enable them to interact with webpages.
All our research is meaningless without evals. So we built & open-sourced Bananalyzer (paper + leaderboard forthcoming) to benchmark codegen agent performance on web data extraction. Over the last few months, we've increased performance from ~30% to >90% on Bananalyzer.
Our codegen agent is currently running in prod with customers spanning govtech, e-commerce, edtech and more. It generates 5K lines of code and extracts >1M rows of data every week. If your business wants to scale up your web data extraction, check us out!
Excited to more broadly share what we've been working on at
@ReworkdAI⚡️ This past year we've fully invested in building the next generation of web agents. They're already live in production today across a handful of companies serving as the simplest way to extract web data at
