FinEval

A benchmark I built to see how well language models actually handle money math — the multi-step, numbers-heavy questions where they tend to sound confident and still get it wrong. Over 1,000 questions, all graded automatically.

FinEval runs models through 1,000+ financial reasoning questions and grades the answers with an LLM-as-a-judge setup, so instead of a single pass/fail number I can see exactly where a model's reasoning falls apart. On top of that there's a Next.js dashboard for digging into the results — comparing models, filtering by question type, and reading the failure cases side by side.