AWS Builder Center
Full Stack Challenge: CraveMap

Full Stack Challenge: CraveMap

CraveMap captures the living, breathing pulse of what neighborhoods are ordering, cooking, and obsessing over at any given moment.

Full Stack Challenge: CraveMap

A map showing the pulse of food craves around your area

I keep a running argument with myself about dinner. Not hunger. Preference. A queue at one stall and an empty counter at the next still does not tell me what people wanted twenty minutes ago. Star ratings sit on restaurants. They age. They talk about a place, not a dish.
CraveMap is that argument turned into a map. People who actually go buy food (night market, high street, the shop downstairs) can drop a craving: a dish name, a mood, a photo, or a pin. Pulse is the heatmap. The feed is what just got posted. Ask CraveMap is a short chat that is only allowed to recommend live posts near you. If Nova invents a ramen stall that nobody posted, the product is lying, so the tool looks up real  DishPost  rows first.
The kitchens are not the users of this build. They are the late buyer. When we scale, I want eateries and restaurants to fund CraveMap because the consumer graph is something they can sell against. The plan is a paid embed, a component or a CDN drop-in, so a vendor can hang the pulse inside an app they already have. QR on the counter. Ads aimed at the dishes and neighborhoods that are actually hot. That layer is conversion. This challenge app is how you collect the signal without faking it.

Vision and what it does

The problem is dish-level demand that dies in group chats. You know a place is “good.” You do not know that three people in Soho just posted kimchi jjigae and that the heat is decaying while you walk.
CraveMap is for the person on the pavement. Sign up with email. Confirm the code and you are signed in. Share location if you want nearby heat; skip and pick a neighborhood. Post a dish. Watch Pulse. Spin Ask if you cannot choose.
Seeded posts are labeled. I would rather a judge see “demo heat” than a heatmap of anonymous fiction.
What it does not do yet is take a vendor’s money. No kitchen dashboard, no QR generator in production, no ad auction. Those are the next product, sitting on top of the same posts and trend scores.
Full stack breakdown
  • Pitch. Stop competing with Google ratings. Compete for “what is this street hungry for tonight,” then sell that to the people cooking.
  • Prototype. A Vite React PWA, MapLibre on OpenFreeMap liberty tiles, a glass tab bar (Pulse, Feed, New, Ask, Profile). I threw out a navy “nightclub” theme because food wanted rice paper and chili, not a cocktail bar.
  • MVP. Amplify Gen 2 in TypeScript: Cognito, AppSync, DynamoDB, S3 photos, Lambdas on streams and a 15 minute decay timer, Bedrock Nova 2 Lite behind Ask. Geohash on the client so nearby queries stay in DynamoDB for now. I know PostGIS would be nicer for radius search. I also know this challenge had a clock.
  • UX. Custom auth screens, not Hosted UI, not the Authenticator widget. Location is a tap, not an ambush: iOS will ignore  geolocate.trigger()  on map load, so Pulse only auto-locates after the user already said yes. Welcome line and avatar on every tab except Profile. Display name is yours to edit; if you leave it blank we parse the email local-part ( alex.smith+news@  becomes Alex).
  • Launch.  amplify.yml  at the repo root,  appRoot: apps/cravemap . Backend phase runs  ampx pipeline-deploy . Frontend is  tsc  and Vite.

How I built it

Cravemap Architecture Design diagram

Sandbox and Hosting tried to own the same global names (Location map, SSM path for the profile table). First Hosting deploy rolled back and left a husk. The fix was namespacing with  AWS_APP_ID  and branch:  cm-${slug}  for Location,  /cravemap/${slug}/user-profile-table  for SSM. Concrete string paths, not  Stack.stackName  tokens, because IAM cannot guess an ARN separator from a token.
Then the backend went green and the frontend failed TypeScript. MapLibre’s  LayerSpecification  union does not put  filter  on background layers.  flatMap  inferred  any[] . I narrowed to  symbol  and stopped spreading the union.
Auth screens existed the whole time. Hosting writes  amplify_outputs.json , so  DEMO_MODE  was false, and  RequireAuth  only gated the demo store. Production opened on Pulse with no session. The gate now waits on Cognito, shows a splash, and auto signs in after the email code.

AWS services / architecture

A signed-in browser talks to Amplify Hosting (CloudFront in front of the static build, S3 under it). Cognito holds users and the  seedAdmins  group. AppSync is the GraphQL API. DynamoDB holds  DishPost ,  TrendScore ,  UserProfile . A stream Lambda bumps neighborhood dish scores. EventBridge fires score decay every 15 minutes. Post-confirmation Lambda writes a profile row (table name from SSM). Photos go to S3 through Amplify Storage. Location Service is a CDK escape hatch on the auth stack (map, place index, geofences), because a nested stack cyclic-depended. Ask CraveMap is an Amplify AI conversation: a handler Lambda, a  findCravingCandidates  tool, Bedrock Nova 2 Lite ( global.amazon.nova-2-lite-v1:0 ). Secrets Manager holds the seed user password. Route 53 points the custom domain. IAM is the glue. CDK is how Gen 2 materializes all of that.
Heat is not a Places “popular times” scrape. A post writes a row. The stream updates a score. The client subscribes. Seeded rows carry  source: SEEDED  so we can purge them.

What I learned

Gen 2 Hosting is not Gen 1 “associate a backend environment.” If you treat the missing association as a failure, you will debug the wrong console.
Global resource names across sandbox and branch stacks will waste a night. Namespace them on day one.
 observeQuery  is pleasant until you forget the session. An unsigned heatmap is a pretty lie.
Browser permissions are a product. A user gesture, HTTPS, a timeout, and a distinct message for deny versus “this WebView will never prompt” are not polish. They are the difference between a map and a shrug.
I also learned I will overfit TypeScript on MapLibre filters and still ship, which is a personality trait, not a methodology.
The next build I want is boring on purpose: the vendor embed, QR, and ads against the same tables. If the consumer pulse is empty, that product has nothing to sell. So the challenge app had to be the consumer side, even though the fundraising story is the kitchen.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article