Pantry Chef

Meal planning with smart grocery shopping. Generate personalized plans, track your pantry inventory, get AI suggestions for cooking with what you have, and checkout at Walmart in one tap

Comment

Pantry Chef: AI-Powered Meal Planning to Grocery Checkout in 60 Seconds

Inspiration

My wife loves cooking, but dietary restrictions turned it into a daily puzzle. Gluten-free, dairy-free, low-FODMAP—every recipe required mental gymnastics: Can I substitute this? Will that ruin the dish? Do I have the ingredients?

When the Ship-a-ton prompt mentioned a chef frustrated with downloading recipes and pantry management, it hit home. This wasn't just my wife's problem—it's every home cook's problem.

The numbers prove it: 78 million Americans have dietary restrictions. Meal planning apps exist, but they treat restrictions as afterthoughts. Recipe sites force you to copy-paste into Notes. Grocery apps make you manually build carts. Nobody connects the dots.

I built Pantry Chef to solve this end-to-end: from meal plan to grocery checkout in under 60 seconds.

What it does

The complete flow:

  1. Generate AI meal plans → Create 3–7 days of recipes with dietary filters (gluten-free, vegan, kosher, low-carb, etc.)
  2. Review & customize → Swap meals, adjust servings for family size, regenerate individual recipes
  3. Smart ingredient merging → AI automatically deduplicates ingredients across recipes ("2 cups flour" + "1 cup flour" = "3 cups flour"), normalizes units, calculates optimal package sizes to minimize waste
  4. One-tap checkout → Send merged shopping list to Walmart with pre-filled cart

Monetization model (dual revenue):

  • Primary: Walmart affiliate fees (we make money when you buy groceries—never gating checkout)
  • Secondary: Subscriptions (Pro $4.99) + consumable credit packs

Credits only gate AI features (real API costs). Recipe usage is completely free—no artificial paywalls. Free users earn credits via sign-up (25 credits) and Walmart checkouts (15 → 10 → 5 diminishing rewards).

Tier features:

  • Free: Meal planning, ingredient merging, Walmart checkout, earn credits
  • Pro ($4.99/mo): 40 credits/month + advanced dietary settings, cookbook scanning, cost tracking, nutrition analytics

How we built it

Tech stack:

  • Frontend: React Native + Expo 54 (cross-platform iOS/Android)
  • Backend: Node.js/Express 5 + Prisma 7 ORM
  • Database: PostgreSQL (Docker containerized)
  • AI: OpenAI GPT-5o for recipe generation and ingredient intelligence
  • Monetization: RevenueCat SDK (subscriptions + consumables)

Architecture highlights:

1. AI Ingredient Merging (95% auto-merge rate)

The hardest problem: recipes are unstructured text. "1 large onion, diced" vs "½ yellow onion" vs "1 medium onion (about 150g)"—same ingredient, three formats.

Two-pass pipeline:

Recipe text → GPT-4 extraction → Structured JSON
{quantity: 2, unit: "cups", item: "all-purpose flour", notes: "sifted"}

Then semantic grouping + unit normalization:

  • Cups → grams for solids (using USDA density tables)
  • tsp/tbsp → ml for liquids
  • "1 large onion" → ~150g equivalent

Result: 95% of ingredients merge automatically. Edge cases (e.g., "diced" vs "sliced whole") flagged for review.

2. Smart Package Calculations

If a recipe needs 3 cups flour but stores sell 5-lb bags, how many bags?

Algorithm: $$\text{packages} = \lceil \frac{\text{total quantity (normalized units)}}{\text{package size}} \rceil$$

Handles fractional packages intelligently (e.g., "buy 1 lime for ¼ cup juice" vs "buy 2 lbs ground beef for 12 oz").

3. RevenueCat Integration

Learned the hard way: concurrent API calls trigger 429 rate limits (error code 7638).

Solution: Multiple useRef guards to serialize operations:

const syncingRef = useRef(false);
const loadingCustomerInfoRef = useRef(false);
const initializingRef = useRef(false);

Run init/refresh sequentially, never with Promise.all(). Unglamorous but critical for reliable purchases.

4. Auth Pattern

JWT tokens stored in AsyncStorage (key: authToken). Helper function getAuthHeaders() centralizes auth logic across API calls.

Challenges we ran into

1. Walmart integration blocked by app store approval

The checkout flow is fully implemented, but I need to publish the app before Walmart grants API credentials for cart creation. Google's identity verification is still pending. This is the reality of shipping in 4 weeks—some gates are outside your control.

2. Ingredient merging refinement

Getting from 70% to 95% auto-merge accuracy required extensive prompt engineering and edge case handling:

  • "1 can (14.5 oz) diced tomatoes" vs "2 cups crushed tomatoes"
  • "Salt to taste" vs "1 tsp kosher salt"
  • "Fresh basil leaves" vs "1 tbsp dried basil"

Each category needed custom normalization rules.

3. Cook-from-pantry AI needs more love

The feature exists (generate recipes from available ingredients), but constraint-based generation is finicky. Sometimes it suggests ingredients you don't have. Needs better prompt engineering and validation.

4. Schema migrations with existing data

Prisma uses db push (not migrate). When adding required columns to tables with data:

  1. Make column nullable first → push
  2. Backfill with SQL
  3. Make required → push again

Otherwise you nuke existing records.

5. The 4-week grind

Evenings and weekends. Docker rebuilds. Wrestling with AsyncStorage. Windows path issues (/c/Users/... vs c:\Users\...). Every feature needed refinement—nothing shipped perfectly on first try.

Accomplishments that we're proud of

1. The complete loop works

Meal plan → ingredient merging → checkout in under 60 seconds. Not "coming soon" or mocked—it actually works (modulo Walmart API approval).

2. 95% auto-merge rate

Users don't notice it until they use a competitor and realize they're manually deduplicating. That's when you know AI is working—it disappears.

3. Sustainable free tier

Affiliate revenue means free users aren't a burden—they're the business model. No ads, no dark patterns, just aligned incentives.

4. Speed as a feature

Meal planning fails when it takes longer than the problem it solves. If it takes 20 minutes to plan dinner, you'll just order takeout. We made it faster than scrolling TikTok recipes.

What we learned

1. AI ingredient merging is harder than it looks

Semantic similarity + unit conversion + package math = surprisingly deep problem. The difference between "works sometimes" and "works 95% of the time" is 10x the effort.

2. RevenueCat quirks matter

SDKs have personalities. Learning to serialize API calls, use multiple ref guards, and avoid concurrent refresh/sync calls was critical for production reliability.

3. External dependencies will block you

Walmart API, Google identity verification, app store approval—you can build the perfect feature and still wait on gatekeepers. Plan accordingly.

4. Dual revenue unlocks better UX

Affiliate revenue means I can make recipe usage free (no artificial credit charges). Subscriptions gate AI costs, not core value. This only works because groceries pay the bills.

5. Shipping beats perfecting

Cook-from-pantry AI isn't perfect. Some ingredients still need manual merge review. But the app is useful today—and that matters more than waiting for 100%.

What's next for Pantry Chef

Immediate (post-competition):

  1. Ship Walmart integration → Pending app store approval and API credentials
  2. Refine cook-from-pantry AI → Better constraint handling, validation that generated recipes only use available ingredients
  3. Improve merge edge cases → Handle "to taste" measurements, fresh vs dried herb conversions

Pro tier features:

  1. Cookbook scanning → OCR + AI to digitize physical cookbooks into structured recipes
  2. Cost tracking → Track spending over time, compare store prices
  3. Nutrition analytics → Macro breakdowns, weekly nutrition summaries

Power tier features:

  1. Pantry tracking → Scan receipts or manually log inventory, get alerts when running low
  2. Family profiles → "Dad hates mushrooms, kid allergic to nuts"—filter recipes accordingly

Long-term (deferred post-competition):

  1. Creator economy → Let users publish recipes, charge credits for usage, revenue share payouts

The real goal: Get this in front of users. The best validation isn't a demo—it's someone saying, "I don't know how I cooked without this."

Built with: React Native (Expo 54), Node.js/Express 5, Prisma 7, PostgreSQL, OpenAI GPT-4, RevenueCat SDK

Built With

Share this project:

Updates