Inspiration

Pantry Talk started from a simple problem: deciding what to cook is often harder than cooking itself. People already have food at home, but they still spend time searching for recipes, buying duplicate ingredients, or letting groceries go to waste. We wanted to build a kitchen companion that starts with what you already have and turns that into something useful.

What it does

Pantry Talk helps users turn their existing groceries into meals they can realistically cook. Users can type or speak ingredients, upload a fridge or grocery photo, and get recipe recommendations ranked by pantry fit, cooking time, and dietary goals. The app also remembers the user’s kitchen across sessions. Recipes can be saved, cooking sessions can be completed and rated, substitutions can be recorded, and Pantry Talk builds a personal cooking history over time. Instead of being a one-off recipe generator, the goal is for it to become a persistent kitchen companion that learns from what users actually cook.

How we built it

Pantry Talk is built with React, TypeScript, TanStack Start, Vite, and Tailwind CSS. We use Supabase for authentication, profiles, persistent pantry data, favorites, collections, and cooking history. Recipe discovery uses a local-first architecture. Pantry Talk first searches and ranks a structured recipe catalog, then can use Linkup to enrich or discover additional live recipes when needed. This reduces unnecessary API calls while keeping recommendations grounded in real recipes and source URLs. We also built structured pantry state, guest-to-account migration, recipe snapshots for durable cooking history, photo-based pantry scanning, voice ingredient input, and a guided cooking mode.

Challenges we ran into

One of the biggest challenges was turning recipe discovery into something genuinely pantry-aware instead of simply returning recipes containing one or two matching ingredients. Ranking recipes required balancing pantry coverage, missing ingredients, cooking time, dietary goals, and usefulness. Persistence also became more complex than expected. We wanted Pantry Talk to work instantly for guests, but also allow users to create an account later without losing their pantry, saved recipes, or cooking history. That required careful local storage, Supabase persistence, deduplication, migration, and row-level security. We also had to make externally sourced recipes reliable enough for a polished product, including handling inconsistent images, changing recipe pages, and preserving enough recipe data so that a user can still revisit something they cooked previously.

Accomplishments that we're proud of

We are especially proud that Pantry Talk has evolved beyond a recipe search interface into a product with real memory. A user can build a kitchen, refresh the page, come back later, sign into an account, and keep that kitchen. They can save recipes, complete cooking sessions, rate meals, record substitutions, and see those experiences reflected back on the recipe later. We also built the recipe system to be local-first rather than LLM-first, keeping common operations faster, cheaper, and more deterministic while reserving external AI and search services for the places where they provide real value.

What we learned

We learned that the most useful AI products are often not about generating more content—they are about maintaining useful context. For Pantry Talk, the important data is not just a prompt asking, “What should I cook?” It is the evolving state of the kitchen: what ingredients are available, what needs to be used soon, what recipes the user saves, what they actually cook, what they substitute, and what they enjoy. We also learned how important deterministic systems are around AI. Search, pantry matching, persistence, deduplication, and ranking work better when they are explicit and testable, with AI layered on top rather than controlling every decision.

What's next for Pantry Talk

The next step is to turn Pantry Talk’s growing memory into deeper personalization. Recommendations will incorporate cooking history, ratings, dietary preferences, pantry state, time constraints, and ingredients that should be used soon. From there, we plan to add meal planning, consolidated grocery lists, and ordering integrations such as Instacart, DoorDash, and Uber Eats—letting Pantry Talk help users decide not only what to cook, but also what to buy when something is missing or what to order when they do not want to cook. Longer term, we see Pantry Talk becoming an intelligent operating layer for the home kitchen: understanding what you have, helping you decide what to eat, guiding you while you cook, reducing waste, and getting more useful every time you use it.

Built With

Share this project:

Updates

Submission history