Developer Builds Tool to Let Postgres Grade AI-Suggested Indexes Automatically

A developer frustrated with unverifiable AI-generated database index suggestions built a small tool that uses Postgres itself to objectively evaluate each recommendation. The tool creates each proposed index inside a transaction, benchmarks the query with EXPLAIN ANALYZE, checks whether the planner actually used the new index, then rolls back — leaving the database unchanged. Timing alone proved insufficient as a metric, since a query could run slightly faster due to noise even when the planner ignored the index entirely. Testing across eight common queries on a 1.5-million-row dataset showed that models typically verified around half to two-thirds of their proposed indexes as genuinely useful, at a total cost of fractions of a cent. The approach highlights a practical gap between AI suggestions that appear plausible and those that deliver measurable, planner-confirmed performance gains.