Technologies We Develop

v0.1.0 (pre-release)

Embeddable Relational Database Engine

A relational database engine written in Rust, shipped as a library, a Postgres-wire-compatible server, a CLI shell and a desktop GUI. Version 0.1.0 is published as a pre-release. It supports a subset of SQL rather than full conformance, and recovery is redo-only without physical checkpoints. Not published to crates.io or npm; build from source.

Release v0.1.0 ↗

v0.1.0

Policy and Security Control Layer for MCP Tool Calls

An open Python library that evaluates policy, risk, budget and delegation rules before an agent or client calls a protected Model Context Protocol tool, and inspects the result afterwards. The package manifest declares version 0.1.0. Development is partway through a documented phase plan: the root distribution packages the core package, while the eleven component packages each carry their own manifest and source. The gateway and control-plane applications described in its documentation are not present in the repository, and no persistence layer is implemented.

v0.2.0 (in-repo, unreleased)

Desktop AI Coding Agent

A desktop coding agent that connects a compatible OpenAI-compatible language model to a local project. It ships a PySide6/Qt interface and a headless CLI, and can read, write, edit and delete files, run approved commands and inspect Git state. Tools declare their own permissions, and a deny-by-default permission layer with a human approval prompt gates every call. Licensed MIT. Limitations: the command policy is an application-level deny list, not an operating-system sandbox, and Git mutations (commit, push, checkout, reset) are deliberately blocked. Version 0.2.0 is declared in the repository but no release or package has been published. Windows is the only platform the GUI has been run on; macOS is listed as planned and BSD is unsupported. An intermittent GUI abort on one CPython version is documented as unresolved.

No validated release

From-Scratch Coding Language Model Research

Open research into training a code-oriented language model from scratch, with its own Transformer implementation (RMSNorm, rotary embeddings, grouped-query attention, SwiGLU), its own BPE tokenizer, training and fine-tuning loops, and local inference plus FastAPI serving for completion and review. Licensed Apache-2.0. The project describes itself as a local prototype. Limitations: no release has been published and no version has been validated. The code corpus committed to the repository is far smaller than the shipped training configuration requests, the committed tokenizer vocabulary does not match the size that configuration declares, and the code-generation evaluation uses a small set of hand-written problems rather than the full HumanEval and MBPP suites it names. Its figures are therefore not comparable to published benchmark results, and no independent model-quality assessment has been carried out.