Two-stage architecture the Nebulab team adopted after rewriting their Cometeer recommendation code. Stage 1 (filtering): start with the full product pool and sequentially apply small filters (matching roast level, caffeine level, inventory, customer dislikes, etc.) to narrow candidates. Stage 2 (scoring): for each surviving product, run an arbitrary number of weighted amplifiers that multiply a running score based on local conditions (votes), then pick the top-scoring product. When no winner exists, the product pool is too small — treated as a supply-chain problem rather than an engineering one. Top competitors are stored as metadata so stakeholders can ask 'if not this one, what would have been picked next?'