Multiversal VenturesOPEN SOURCE
What we publish in the open
Two problems sit under every offering we have: reading documents reliably, and keeping sensitive data on the machine. A regulated buyer should not have to take our word for how either is done. So the parts that can be inspected, are.
01
OpenReading
The unified interface for every reading.
Every OCR and document provider fails somewhere. OpenReading is one consistent interface and one response schema across fifteen-plus backends, from Tesseract and PyMuPDF to Textract and Google Document AI, so switching or combining providers is configuration, not a rewrite. Nothing routes through our infrastructure.
- Parse
- Read a document through any backend, into one normalized schema. Files, folders and globs.
- Compare
- Run several providers on the same document and see exactly where their readings agree and where they diverge.
- Strategy
- Cascades and races with quality gates, written in YAML and testable offline. When one provider fails, route around it.
- Interfaces
- CLI, Python library, HTTP server.
- License
- Apache-2.0 core.
02
Cuttlefish
Native tooling for AI agents.
A coding agent hands off a job, summarize this, classify that, extract these fields, and Cuttlefish runs it locally against a local model and returns a structured result. Sensitive files never go to a frontier model, and routine work stops costing frontier prices.
- Data locality
- Files stay on the machine. Processing blocks receive handles, not contents.
- Sandboxed execution
- Blocks are WebAssembly, isolated, with capability restrictions checked at compile time and at run time. A job declares which model, which files and which network it may touch.
- Pipelines
- Blocks connect into dataflow graphs. Text, images, and PDFs with or without a text layer.
- Cost
- Routine work runs on small local models instead of large remote ones.
- License
- Apache-2.0 or MIT, at your option.
03
Where these show up in our work
Document reading is the first step of nearly every studio engagement, and of Fenero and Brook. Local, sandboxed model execution is how an agent workforce andCartographer can work inside a client’s boundary without moving their data. The research behind both comes out ofour lab.