MealMate

MealMemory
MealPrep

Inspiration

Every evening my family has the same conversation — "what should we cook?" With a diabetic family member, a vegetarian spouse, and twin daughters with strong preferences, it's not a simple question. It's a daily optimisation problem that takes 20-30 minutes. We built MealMate to end that conversation for good.

What it does

MealMate is an AI agent that handles family meal planning end-to-end. Snap a photo of your fridge — it identifies ingredients using computer vision. Tell it about your family once — it remembers dietary restrictions, preferences, and feedback permanently. It generates personalised weekly meal plans, school lunchboxes (no heating, finger-friendly, 5 mins to pack), and shopping lists that exclude what you already have. Every suggestion respects everyone's needs simultaneously.

How we built it

Built with AWS Strands Agents SDK and Amazon Bedrock (Claude Sonnet). One agent with five tools: vision (fridge scanning via Bedrock Converse API), meal planner, shopping list builder, pantry tracker, and memory search/add. Session persistence via FileSessionManager for conversation history. Long-term memory via TestMemoryStore for family preferences that survive across sessions. Streamlit frontend for the UI.

Challenges we ran into

Getting the memory layer to work reliably was the hardest part. The agent would save preferences but fail to retrieve them — the keyword-based search didn't match well against stored entries. We had to tune the system prompt to instruct the agent on which search terms to use. The Bedrock vision API also needed the correct inference profile ID format rather than the direct model ID — a subtle AWS configuration detail that took debugging.

Accomplishments that we're proud of

The fridge-to-dinner flow works end-to-end. Upload a photo, the agent identifies ingredients, recalls your family's dietary needs from memory, and generates meals that work for a diabetic, a vegetarian, and two picky kids — all in one response, with cooking times and a shopping list. It feels like a product, not a demo.

What we learned

Prompt engineering is 70% of the work. The agent's intelligence comes from how you instruct it — what search terms to use for memory, when to prioritise fridge ingredients over preferences, how to format lunchbox constraints. The framework and tools are straightforward. Getting the agent to behave consistently is the real engineering challenge.

What's next for MealMate

Multi-image support for scanning the fridge, pantry, and spice rack separately. Feedback learning — "kids loved the tacos" or "soup was too spicy" shapes future suggestions. Integration with grocery delivery APIs for one-click shopping. Recipe card generation with step-by-step cooking instructions. Deployment to AWS AgentCore for production hosting.