hey heres a little thing i built for working on long running ml projects. with the whole karpathy/autresearch thing blowing up thought this would be worth sharing. theres much inefficiency in remote training jobs, also if stuff fails to converge or something breaks you end up burning much compute for nothing. so rlclaw is your research companion that can ping you on discord for status updates, you can steer it, has checkpointing. and mega bonus i reversed the colab extension and built a little tool so you can grab the colab auth token as if you are the vscode extension, this allows using colab pro+ with rlclaw. rlclaw lives on my workstation with a 5070ti but if its hungry it can use colab for a kernel on a t4 or a100 box
FYI- I built a simple tool that lets you OCR PDFs entirely in your browser using Tesseract.js and PDF.js. No server required - all processing happens client-side. this tool creates a text transparent layer that preserves original document layout using bounding box coordinates like most ocr tools.
