Your idea works.
It just doesn't exist yet.

A reminder app for prayer times. The user gets a notification before the relevant window closes wherever they happen to be, without setting anything up again each day. What you see from the outside is a notification arriving on time. What it takes is a send engine that fires on the right minute, every day, for every location.

• Prayer times come from the Hebcal zmanim API. Notifications go out through Firebase Cloud Functions, Firestore and FCM.
• There is no task queue. A function runs once a minute and checks whether a Firestore document exists for the current minute. Documents are keyed hour_minute, so 13:47 is the document 13_47. If it is not there, the run does nothing.
• Users are grouped by rounded latitude and longitude, so the zmanim API is called once per location per day instead of once per user. Each run also writes tomorrow's document for that group, so the schedule advances one day at a time instead of being precomputed.
• Two separate and independent ways to stop receiving: a permanent off switch and a same-day snooze. Both get checked before every send.

An ordering site for the Ra'anana branch: menu, deals, cart, payment and delivery. From the outside it looks like a menu. The real work starts the moment a customer hits checkout.

• Built in Vite and React on top of Base44, with card payments through Cardcom.
• Orders are handed to the branch's Aviv POS, so nobody retypes them at the register.
• The order of operations around money: a pending order is created first, then payment runs at Cardcom, and only after that come the status update, the emails and the POS handoff.
• Every delivery zone carries its own delivery fee and minimum order, and those two values decide whether an address can complete an order at all.
• A deal is not a discounted product. It can hold several products with its own option choices, so the cart keeps the deal's structure instead of flattening it into one line.

A two-sided recruitment workspace: candidates on one side, roles and recruiters on the other, and the matching between them. Public plans start at $29 a month.

• The data sits in PostgreSQL, with the platform built on Base44.
• Embeddings use OpenAI's text-embedding-3-small, computed over structured statement parts rather than one text blob per record.
• That is what makes a result explainable. Cosine similarity gives a proximity score, and a score on its own is not a decision — a candidate can look close to a role and still fail a hard requirement, so the score passes through simple pass/fail business rules.
• The matching screens read from snapshot tables rebuilt once a day, instead of scoring every candidate against every job each time a page opens.
• Two audiences needing different things out of the same data is what made this a six-month project instead of a one-month one.