Open Source

Tools and research we share

We build AI systems for clients, and some of what we learn along the way is useful to everyone. These are the models, tools, and research collaborations we've released, with the code and docs to try them.

Pento on GitHub

With Qdrant

Bilingual sparse retrieval

miniCOIL EN-ES

A sparse retriever that matches English and Spanish in a single index, with no translation step. It's the next step in Qdrant's miniCOIL line and its first release beyond English: Pento did the development end to end, with Qdrant as technical advisor.

2,398 bilingual concepts in about 30 MB on top of its encoder
Ranks Spanish results better than a BM25 configured for Spanish
Falls back to plain BM25 for words outside its concepts
Research preview under CC BY-NC 4.0, for research and evaluation

Language: Python
License: CC BY-NC 4.0

Try it

$ git clone https://github.com/pentoai/minicoil-en-es

GitHub Hugging Face Qdrant write-up Our blog post

With alphaXiv

Research papers inside your coding agent

alphaXiv MCP

An MCP server that connects alphaXiv's research corpus to Claude Code, Cursor, VS Code, Zed, and other MCP clients, so your agent can search papers, read them, and inspect their code without leaving the editor. Pento is one of alphaXiv's collaborating organizations.

21 tools for papers, researchers, and your library
Search and read papers, query PDFs, and read a paper's GitHub repository
OAuth sign-in, or an API key for headless agents

Try it

$ claude mcp add --transport http alphaxiv https://api.alphaxiv.org/mcp/v1

Docs About alphaXiv Our blog post

Autonomous ML experimentation agent

ml-ralph

An ML engineering agent with a terminal UI that runs on Claude Code. You define the goals in a PRD; it plans experiments, runs them, tracks metrics, and records what it learns, trying to break its own results before it trusts them.

Language: TypeScript
License: MIT

Try it

$ bunx @pentoai/ml-ralph

GitHub npm

Embedding experiment tracking

Vectory

Track and compare embedding versions: which data and model produced each one, and how their vector spaces compare on your metrics. Works from the CLI, Python, or a GUI.

Language: Python
License: MIT

Try it

$ pip install vectory

GitHub Our blog post

Face-recognition video timelines

Streamlit Terran Timeline

A Streamlit component that builds a face-recognition timeline for a video, showing who appears on screen and when. Built on the Terran human-perception library.

Language: Python
License: MIT

Try it

$ pip install streamlit-terran-timeline

GitHub PyPI

Have a hard search or ML problem?

Book a 30-minute scoping call Schedule a meeting

The engineers behind these projects build production AI systems for clients. Tell us what you're working on.