🍲 Khana AI — Zero-Waste 30-Day Smart Meal & Grocery Agent

💡 Inspiration

Every day, millions of households waste 30 to 45 minutes arguing over two exhausting questions:

1. "What should we cook today?"
2. "What and how much should we buy from the market?"

This daily friction leads to repetitive meals, nutritional deficits, and massive food waste in refrigerators when perishable produce spoils. We created Khana AI — an autonomous AI Agent built with the AWS Strands Agents SDK and Amazon Bedrock — to take the chore out of daily meals, eliminate food waste, scale market grocery buying in exact kg & gm for any household size, and guarantee 100% USDA balanced nutrition.

🍱 What It Does

• 30-Day 90-Meal Matrix: Generates 90 distinct, location-tailored meals (Breakfast, Lunch, Dinner) across 30 days based on selected Country and State/Region (e.g., Gujarat, Punjab, South India, Maharashtra, Bengal).
• Strict Dietary Guardrails: 🟢 Vegetarian and 🥬 Vegan modes enforce 100% zero-contamination guardrails — automatically stripping meat, seafood, dairy, or honey across all 90 meals.
• Household Member Portion Scaling: Dynamically computes household scale factors (e.g., 2 Adults + 1 Child = 2.6x scale) and displays exact ingredient weights in grams (gm) on recipe cards.
• Smart Market Grocery Checklist (kg & gm): Canonicalizes produce into clean department categories (Grains, Fresh Veggies, Dairy, Oils, Spices) and calculates total market purchase weights in kg and gm for any custom duration (Today, Tomorrow, Tomorrow + 2 Days, Weekly, or 30-Day Bulk).
• USDA Nutrition Benchmark Dashboard: Real-time RDA compliance gauges comparing daily totals against official USDA nutrition standards for Calories, Protein, Iron, Calcium, and Vitamin C.
• Strands AI Agent Playground: Interactive agent interface powered by @tool decorated functions with live execution trace visualization.

⚙️ How We Built It

• AWS Strands Agents SDK (strands-agents): Built the core AI Agent in Python with decorated Strands tools:
  • @tool usda_nutrition_analyzer_tool: Fetches nutritional profiles from USDA FoodData Central API.
  • @tool grocery_market_scaler_tool: Scales recipe base ingredients by family size and shopping frequency into exact kg and gm.
  • @tool strict_diet_guardrail_tool: Intercepts and blocks non-veg or non-vegan ingredients.
  • @tool Regional_Cuisine_Adapter: Dynamically adapts 30-day plans to authentic regional specialties.
• Amazon Bedrock Foundation Models: Integrated via boto3 Bedrock Runtime (us.amazon.nova-pro-v1:0 and Claude 3.5 Sonnet).
• Amazon Bedrock AgentCore Deployment: Containerized with Docker and agentcore.json for serverless AWS AgentCore deployment.
• Web Frontend: Built a zero-dependency, ultra-fast single-page web app deployed live on GitHub Pages.

🚧 Challenges We Faced

1. Ingredient Duplication & Canonicalization: Consolidating variations like "Whole Wheat Atta", "Wheat Flour", and "Roti Dough" into a single clean market shopping entry without duplicating entries.
2. Relative Day Calculations: Dynamically mapping shopping lists to relative user timelines (Day N Today, Day N+1 Tomorrow, Day N+1 to N+3).
3. Dietary Guardrails: Ensuring strict 100% vegetarian/vegan filtering across all 90 generated meals.

🎓 What We Learned

• How to structure production-ready agentic architectures using the AWS Strands Agents SDK.
• Extending decorated Strands @tool functions to handle complex JSON transformations and guardrail verification.
• Integrating USDA FoodData Central APIs into real-time LLM agent reasoning loops.

🔮 What's Next for Khana AI

• Smart Fridge Camera Integration: Scanning refrigerator contents to automatically adjust the grocery buying list based on existing inventory.
• Local Grocery Delivery APIs: One-click checkout to send the generated market list directly to local grocery delivery platforms.

Built With

• amazon-bedrock
• amazon-web-services
• bedrock-agentcore
• boto3
• github
• javascript
• pydantic
• python
• strands-agents-sdk
• tailwind-css
• usda-api