Inspiration
As university students who enjoy cooking our own meals and meal-prepping for the week, we often find ourselves facing the same dilemma: opening the fridge, staring at its contents, and spending minutes trying to figure out what to make for dinner. Somehow, we always seem to end up with the same basic chicken-and-rice meal, even when we have plenty of other ingredients available. We realized that the problem wasn't necessarily a lack of food, but a lack of inspiration. That's how SmartPantry was born: an idea to make everyday cooking easier, more creative, and more enjoyable.
What it does
You take a few photos of whatever is in your fridge or pantry, and SmartPantry figures out what you have on hand. It pins little interactive tags directly over your food in the photo so you can double-check the quantities. From there, it recommends actual recipes ranked by what's about to go bad first and what matches your dietary preferences or protein goals. Once you pick a recipe, there’s a completely hands-free cook mode: you just say the keyword "voiskey" followed by a command (like "voiskey, next step" or "voiskey, start a timer for 5 minutes"), and it reads instructions out loud and sets timers so you never have to smudge your phone screen with messy hands. When you're done, you hit "I cooked this," and it automatically deducts the ingredients from your pantry so your inventory stays up to date.
How we built it
We started by brainstorming everything in a Google Doc, outlining the exact screens and edge cases we wanted. Then we created an AGENT.md spec file to keep our architecture, rules, tech stack, and design tokens organized. From there, we used Google Gemini to create a markdown file for each feature, describing all the requirements an AI agent would need to build the application. Throughout the process we reviewed the code generated and instrucuted more steps.
Challenges we ran into
One of our main challenges was getting consistently formatted outputs from Gemini. Its responses would sometimes deviate from our expected structure, making the generated specifications harder to use reliably. To address this, we experimented with different system prompts and iteratively refined our instructions to produce more structured and consistent results.
Accomplishments that we're proud of
We were really proud of the voice guided recipe instructions.
Log in or sign up for Devpost to join the conversation.