FoodieMama.

The problem

Restaurant discovery is broken in a specific way. Google and Yelp own breadth but are drowning in fake reviews and reduce a restaurant to a star average; short-video drives an enormous amount of real discovery — Google's own data says ~40% of 18-to-24-year-olds use TikTok or Instagram, not Search or Maps, to find lunch[2] — but those surfaces are unsearchable, unsaveable, and unmonetisable for the viewer. You see a great spot in a Reel and have no way to find it again.

FoodieMama exists in that gap: capture the trust of creator/friend curation, make it searchable on a map, and put an AI concierge on top that recommends named places instead of a crowd average.

Market & opportunity

There's no credible standalone "restaurant discovery app" market figure — discovery sits across the US restaurant industry (~$1.5T in sales[1]), the food-tech market (~$210–260B), and reservations. What's measurable is the behaviour shift. DoorDash's 2025 survey found 52% of diners are comfortable with AI recommending where and what to eat (60% of Millennials), while friends-and-family recommendations (53%) outrank influencers as a dining input[3]. And the social-curation thesis has a live proof point: Beli logged 30M reviews in 2024, up from 13M in 2023, ~80% of users under 35, earning a Forbes Cloud 100 Rising Star nod[4][5].

The incumbents see it too: Google is rolling Gemini into Maps for conversational "vibe" queries, OpenTable shipped an AI Concierge, and Tripadvisor partnered with Perplexity[6]. As AI assistants compress choice to 3–5 named brands per query, being the recommended one — through trusted curation rather than a star average — becomes the whole game.

Who it's for

The under-35 diner (and the food-traveller curating for them) who discovers on social but wants it organised: a map they can browse, lists they can save, and an assistant that knows the curated set. India is an explicit second market — street food is formalising fast, and the country is set to be the 3rd-largest food-services market by 2028, with the organised segment growing ~13% CAGR[7].

Constraints

• The AI concierge can't hallucinate restaurants. A food recommender that invents a place is worse than useless — every recommendation has to be grounded in real, curated places.
• The Groq key never ships client-side, and free AI usage has to be capped or it becomes an open LLM proxy.
• Maps with hundreds of POIs must stay smooth on a phone — clustering, not a thousand pins.
• Two-sided monetisation from the start. Consumer Pro and vendor/brand billing are different payment rails (RevenueCat vs Stripe) that both have to be wired correctly and reconciled.

Architecture & what I built

The grounded AI concierge (RAG, not vibes)

The concierge runs on Groq (Llama 3.3 70B) inside a Vercel Function, never the client. The key move is places-context injection: the relevant curated/seeded places are passed in and built into a numbered context block in the system prompt, so the model recommends from a real, named set rather than its training memory. A guardrail module validates input, enforces an IP-level rate limit plus a per-user FREE_DAILY_CAP (Pro users skip it), and keeps the system prompt server-side.

Map-first discovery + contribution loop

Discovery is a clustered react-native-maps+ Google Maps surface with location-based filtering, backed by Supabase (places, bookmarks, comments, device tokens). Users save spots to lists and contribute photos and reviews; a Supabase seed pipeline pre-populates a curator's places so the app is never empty on first open — solving the cold-start problem that kills discovery apps.

Two-sided payments

• RevenueCat handles consumer Pro tiers (monthly / annual) with the paywall and entitlements.
• Stripe handles the vendor side — brand-deal and vendor-listing checkout (Bronze/Silver/Gold), with webhooks syncing subscription and charge state back into Supabase, and a RevenueCat webhook reconciling consumer events into the same ledger.
• Auth is Google OAuth + Apple Sign-In across web, iOS, and Android.

Trade-offs

• Curated/seeded set over open crowd-sourcing.Starting from a curator's places means narrower coverage early, but it guarantees quality and a non-empty first session — and it's what makes the AI concierge trustworthy. Breadth comes later via contributions.
• Groq over a frontier model for the concierge. The recommendation task is grounded by RAG, so latency and cost matter more than raw reasoning — Groq wins that trade for a chat surface people use casually.
• Two payment systems over one. RevenueCat + Stripe is more surface to maintain than a single rail, but consumer subscriptions and vendor invoicing are genuinely different problems with different best-in-class tools.

Goals & what's next

Outcome

FoodieMama is live on the App Store and Play Store, shipping the grounded-AI-concierge + creator-curation + two-sided-revenue thesis as a working product. The bet: in a market where AI search is collapsing dining choice to a few names and anonymous reviews are losing trust, curation is the moat— and the app that makes a trusted person's taste searchable, mappable, and monetisable wins the under-35 diner the incumbents are losing to social.