Component catalogue

MarkItDown (microsoft/markitdown) is a lightweight Python utility that converts many file types to Markdown for LLMs and text-analysis pipelines. README positions it near textract, but focused on preserving structure (headings, lists, tables, links) as Markdown for models rather than high-fidelity human publishing.

turbopuffer is a managed, commercial search engine from turbopuffer Inc. (Canadian, founded 2023). It is object-storage-first: Amazon S3 / Google Cloud Storage / Azure Blob is the source of truth; NVMe and memory are caches. It does approximate vector search (SPFresh), BM25 full-text, filters, and hybrid fusion.

pdf-inspector (Firecrawl, MIT) — pure Rust library (+ Python / Node / WASM bindings) that: 1. Classifies PDFs as text-based / scanned / image-based / mixed in ~10–50ms 2. Extracts position-aware text and clean Markdown (headings, lists, tables, code) without OCR 3.

Generative Loaders is an MIT-licensed React package and live gallery for loading states specific to generative interfaces: streaming-text reveals, compact inline activity indicators, and image-generation placeholders. Install with npm install generative-loaders; it supports React 18+ and uses Framer Motion.

MIT-licensed Node CLI + agent skill by Peter Yang. It opens a target HTML/Markdown file or localhost route in a local browser review UI. The reviewer can directly change text/basic formatting, resize images, delete elements, and attach comments to selected text or blocks.

Exact author text: > We are open sourcing cn, a new engine for Tailwind class merging and conflict resolution. It is a drop-in replacement for tailwind-merge and clsx. > > It works with all frameworks, runs everywhere and is 30x faster. > > Brought to you by @aidenybai and me.

Laya = Apache-2.0, non-autoregressive System One decision model family: given a state (text / email / ticket / JSON) and typed questions (choice / score / noul), returns typed answers with probabilities in one forward pass. Trained with RLCD (reinforcement learning against strictly proper scoring rules).

Needle 3 is Cactus Compute’s on-device “automation foundation model”: a laddered / sliceable tiny model (post: 8–29 MB CQ2-bit binary; 25–121M params; depths 2–20 layers from one weight set) built on their Simple Attention Networks, aimed at tool calls, structured extraction, and local text embeddings — not open-ended chat.