CookFlow is a multi-agent AI kitchen assistant that helps households plan meals around what they already have, reduce food
waste, and eat well without the weekly stress of starting from scratch. Built as part of Aggregate Intellect Bootcamp, v2 was
shaped by 12 in-depth user interviews conducted across Canada.
Objective
Version 1 set out to reduce meal-planning stress through structured batch-cooking. Version 2 reframes the core problem: most
users don't start with a recipe, they start with what's already in their fridge. CookFlow v2 meets users where they are, with
ingredient-first planning, faster responses, and memory that persists across sessions.
What Changed in v2:
Twelve interviews revealed patterns that reshaped the product:
Ingredient-first search: "What can I make with what I have?" is now the primary entry point, not recipe browsing
Constraint hierarchy: Hard constraints (allergies, medical conditions) are enforced strictly; preferences are treated as
optimizable, not blockers
Faster recipe retrieval: Reduced latency in the Recipe Finder agent; fallback suggestions when constraints conflict
Persistent portion memory: Household size and preferences are remembered across sessions
Anti-waste planning: Leftover-aware suggestions surface before produce spoils
Cultural cuisine support: Ingredient substitution guidance for multicultural and newcomer households
These changes led to a more intuitive, personalized, and waste-conscious meal planning experience.
Solution Architecture:
Version 2 restructured the system from 6 agents to a leaner, more reliable design based on a formal agent vs. service classification:
Component
Type
Role
Root Agent
Agent
Orchestration, clarification loop, routing, fallback decisions, final assembly
Python ingredient consolidation and quantity scaling (no LLM arithmetic)
Recipe DB Fallback
Utility
Curated local recipe DB when live search returns empty
User Research:
12 in-depth interviews conducted across Ontario and Alberta, covering households from 1–5+ people: families with young
children, multicultural/newcomer households, individuals managing medical dietary conditions (hypothyroidism, celiac disease,
diabetes), and large families with teens.
Key findings that drove v2 design decisions:
10/12 users plan ingredient-first, not recipe-first
8/12 named food waste as a primary emotional pain
12/12 cited time as their top constraint
6/12 are multicultural households with unmet ingredient-sourcing needs
Condition-specific diets (hypothyroidism, Down syndrome + autism) are completely ignored by every existing tool
Tools
Gemini API
Google ADK
Python
FastAPI
GCP Cloud Run
Firestore
Achievements
Redesigned architecture from 6 agents to 2 agents + stateless Python utilities, reducing LLM token usage by 40–60%.
Conducted 12 in-depth user interviews across Canada to validate and reshape the product.
Hardened allergen safety with deterministic Python enforcement on every request.
Improved average evaluation score from 1/5 (v1) to 3.67/5 across 8 structured test cases.
Built ingredient-first planning mode as the primary entry point, directly from user research.
Deployed a working agentic system accessible to real users on GCP Cloud Run.