Wisent

Wisent is an AI company whose technology is based on representation engineering: reading and modifying the internal representations ("concept vectors" in a model's latent space) of large language models in order to control their behavior. The company positions this as an alternative to treating models as black boxes, claiming it can reduce hallucinations and malicious content, improve coding and math ability, and personalize model behavior without lengthy fine-tuning [0][1][4].

Its stated delivery surface is a developer package with a simple API and SDK, designed to work with existing models and to be deployed either as a cloud API or on-premise for sensitive applications; the site also advertises a demo booking flow and an iOS app [0][1]. Wisent's public engineering output is organized under the wisent-ai GitHub organization, which lists 53 repositories across Rust, Python, Swift, TypeScript, JavaScript, C++ and Shell [3]. The most prominent is Ster, described in its README as a native Rust toolkit for representation reading and activation steering in open-weight language models; Wisent is identified there as the company that builds Ster [2]. Other repositories include uncensorbench, singularity and singularity-desktop (described as a native macOS interface for an autonomous "Singularity digital being"), singularity-landing (a public product website for Singularity), stado, brama, las, wisent-python, wisent-extractors and wisent-evaluators [3].

A third-party explainer article describes Wisent as a platform and toolset that operationalizes academic representation-engineering research into accessible tools, naming an open-source Python package called wisent-guard alongside a commercial API [4].