With 10 SQL queries across 171 rows, the compute burden is dominated by LLM synthesis rather than data retrieval—SQL execution on this dataset scale is negligible. This workload is suitable for scheduled runs on a daily or weekly cadence, though on-demand execution remains viable given the modest query volume. To reduce future compute cost, prioritize filtering at the SQL layer (e.g., funding stage or geography constraints) before passing rows to LLM processing, since the bottleneck lies in synthesis throughput rather than database operations.