Participated in public demo 2601 with CookFlow

The problem [00:00] Sandra framed CookFlow around the everyday exhaustion of deciding what to cook, especially in households juggling different food cultures, allergies, or generational dietary needs. In her own interviews, 8 of 12 people cited food waste as one of their biggest frustrations, driven less by cost than by guilt. What they built [01:05] CookFlow is a multi-agent AI kitchen assistant, built with Google ADK and Gemini and deployed on GCP Cloud Run, that plans a household's week, recipes, grocery list, and a batch-cooking schedule, from stated preferences and constraints. [02:35] After user testing revealed her original architecture (six agents enforcing every constraint equally, with no fallback when search failed) missed how people actually think, 10 of 12 interviewees start from what's in the fridge, not from a recipe, she redesigned it down to three agents with a constraint hierarchy and two entry points: a structured form and a conversational chat, both supporting recipe-first or fridge-first planning. What happened [05:05] After the redesign, her hardest baseline test case went from 1/5 to 5/5, allergen constraints were never violated across any test case, and the system now flags unmet constraints explicitly instead of silently returning nothing. The ask [05:30] Sandra is asking busy families to try CookFlow, inviting anyone who sees a business opportunity to talk, and open to connecting with anyone looking for someone who can take an AI product from user insight to a deployed system.