Paredown is an evaluation layer for AI-generated advertising. Modern agents can produce hundreds of ad creatives in an afternoon, but nobody can tell which of them deserve real budget — so teams either waste money testing everything, or trust gut feel. Paredown fixes that with measurement instead of vibes.
The product ranks AI-generated Meta ad creatives before any spend happens. Under the hood
it is deliberately boring: image and copy embeddings feed a gradient-boosted ranker, every
claim is validated with leakage-safe chronological backtests against your own historical
Meta export (never splitting ads from the same campaign across train and test), baselines
are always reported alongside the model, and confidence intervals come standard. When there
is not enough data to answer honestly, Paredown returns insufficient_data rather
than inventing a score.
Paredown is built agent-first: it ships as a hosted MCP server that Claude, Codex, Cursor, or any MCP-capable assistant can use directly, plus a local CLI for teams who prefer terminals. The company is self-funded and small; we run on the same tools you do.