Skip to content
Back to work

Evolutionary prompt optimization · Google hackathon 2026

Granum

Optimizing appeal-letter prompts by hand doesn't scale or stay reproducible.

View source

What I built

An evolutionary optimization engine modeling immune-system dynamics — structured mutation operators, an LLM-as-judge tournament with median-of-3 sampling, citation-based negative selection, clonal champion memory, and antigen-drift detection — architected around a pluggable Gemini/Vertex judge backend with a Next.js audit dashboard for reproducible head-to-head prompt evaluation.

Stack

PythonGeminiADKNext.js

Status

Engine shipped (159 tests); built for a Google hackathon (Jun 2026).