Libraries

A retrieval engine that reasons over document structure instead of embedding it. Parse a document into a tree, let a model navigate the tree, return whole sections with citations.

The TypeScript SDK. Structure-preserving retrieval without chunking, without a vector database, with every choice traceable.

The Python SDK for the same engine. Knowing by reasoning, not vectors.

LiteLLM for Go. One provider-agnostic client over Anthropic, OpenAI and Gemini, with a router, fallback, cost tracking and capability flags. It is the gateway the engine depends on, extracted so anything in Go can use it.

A Go-native port of Python’s pdfplumber. Go had libraries that render PDFs, edit metadata or pull bag-of-words text, and none that hand you a positioned object model you can run table detection over. Now it does.

Collective problem-solving memory for coding agents. Context7 gives an agent the docs; this gives it what the docs do not cover. Every uncommon error an agent solves used to vanish with the session.

A faithful TypeScript port of youtube-transcript-api. It calls the same internal player endpoint rather than scraping the page, so it does not break when the DOM changes. Node, Bun and Deno.

A community Python SDK for YarnGPT, text to speech in Nigerian accents. Typed, async, and with a CLI, because the API had no client.

120 metaheuristic optimisation algorithms as drop-in torch.optim.Optimizer subclasses — 60 for training weights without gradients, 60 for hyperparameter search — with population fitness evaluated in one batched pass on the GPU.