{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/f469e6eac9c246f6b5de68e52429a275\" frameborder=\"0\" width=\"1660\" height=\"1245\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1245,"width":1660,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1245,"thumbnail_width":1660,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/f469e6eac9c246f6b5de68e52429a275-0ad27b2f23595b5a.gif","duration":145.82,"title":"Fireworks AI: Receipts Demo","description":"This Loom explains how to reduce the cost of AI support triage by switching from a paid model to cheaper open models with a controlled evaluation process. The creator claims they saved a pharma company $12 million a year by proving 63 applications were not worth paying for, and applies the same math to prompts. The proposed tool runs real prompts on both the current model and Fireworks open models, grades outputs deterministically using code instead of another AI, and shows a single verdict per workload. It routes easy questions to the open model and keeps hard ones on GPD, aiming for 86 percent cheaper results with switch rules based on at least 20 percent savings and zero regressions, and reruns on every pull request to prevent prompt changes from breaking things."}