Projects

Bragi turns a document into a bgraph—one structured, addressable graph of its sections, paragraphs, and content, with the coordinates to point back at the page each piece came from. Built for retrieval-augmented generation.

Bragi understands document structure—headings, sections, paragraphs—based on visual and semantic cues, not arbitrary chunking. Node ids are derived from content and position, so a reparse after an edit changes only the edited node, and every node maps to exact PDF coordinates for grounding answers back on the original page.

A frontend runtime born from a simple observation: frontend state management is a distributed systems problem. By accepting this reality rather than hiding from it, we unlock elegant solutions that feel natural rather than forced.

State lives orthogonal to DOM. Ownership and borrowing prevent synchronization bugs. A game-loop renderer provides eventual consistency. Convention over configuration keeps the framework at a small bundle size.

A memory layer for AI collaborators—a knowledge graph layered over content-addressed, structurally parsed input. Where most AI memory is opaque text run through an extraction pass and vector search, Kvasir keeps the topology: documents, conversations, and code interlink as structure, not as inferred similarity.

Bragi reads, Kvasir remembers. Every artifact exists at multiple resolutions through layered summaries, giving logarithmic retrieval depth—vector search becomes optional, not load-bearing.

An append-only, content-addressed, encrypted store of things that have been committed—documents, nodes, versions, attestations. Named for the Norn who governs what has been: what falls into the well cannot be changed.

Urd guarantees three properties at once that are hard to get together: integrity (content addressed by its hash), provenance (signed, chained commits), and privacy (encrypted at rest—the store never holds the key). The storage floor the rest of the stack builds on.