AI safety evaluation lab
A hands-on harness for black-box LLM safety evaluations: a scenario runs against a target model, the response flows through an evaluator DAG combining LLM-as-judge with semantic analysis, and scores are aggregated into a report streamed live to the browser over SSE.
It also includes a "semantic space lab" for exploring embedding geometry against a locally hosted model — cosine similarity, 2D PCA projections, and vector arithmetic (A − B + C) — useful for building intuition about what embedders actually encode.
Project information
- Technology used: .NET 10, ASP.NET Core, React + TypeScript, LM Studio (local models), Server-Sent Events
- Project goal: Understand and measure LLM safety behavior hands-on
- Date: 2026