WTM? | Devpost

Inspiration

People see cool places on TikTok or Instagram, save them, then forget. Or they want to go out but don't want to scroll through videos to find the name or address.
TikTok and Instagram show what's trending. We pull places from those posts, verify they exist, and show them on a map so you can see what's popular near you.

What It Does

WTM is an iPhone app that shows a map with pins for trending places ranked by a trend score from views, likes, comments, and recency. Tap a pin to see the name, address, rating, category, photos, a short description, and the posts that made it trend.
You can search with phrases like "chill coffee spots" or "hidden gems." The app uses AI to interpret your query. You can filter by category (restaurants, coffee, etc.). The app is read-only, that means no accounts, and it uses our public API.

How We Built It

Data Collection

Daily batch jobs scrape TikTok and Instagram via Apify for hashtags like Dallas, dtx, uptowndallas.
Each post: platform, url, caption, likes, views, comments, date, thumbnail.
Jobs run automatically at 6 AM UTC via Celery Beat; can also be triggered manually. Configurable via environment variables.

Finding Locations

For each post, we ask NVIDIA Nemotron which real-world location it’s about.
If none is found, we skip; if it returns something like "Local Good Coffee in Dallas", we proceed.

Verifying Places

Google Places confirms the place exists and provides official name, address, coordinates, category, and rating.

Creating Descriptions

After creating a location, Nemotron generates a short description from Google Places reviews and post captions (2–3 sentences).
Runs automatically in the background.

Storing Data

Stored in PostgreSQL with PostGIS for location queries.
locations: name, address, coordinates, rating, category, images, trend score, description.
media: one TikTok/IG post linked to that location.
On new post insert, recompute a 0–100 trend score:
- TikTok emphasizes views, Instagram emphasizes likes
- Consider engagement speed and recency
- Boost for lots of recent activity

Public API

GET /api/locations/nearby - Returns places within a radius, ordered by trend score then distance.
GET /api/locations/search - Natural-language search using Nemotron to interpret queries (keywords, categories, rating ranges). Falls back to basic text search if the LLM fails.
GET /api/locations/{id} - Full place details including description, preview images, recent posts with links, and reviews.

iPhone App

Map-first screen using MapKit with pins and a bottom sheet of nearby places.
Top filter chips (Restaurants, Coffee, etc.) filter client-side using backend categories.
Detail view: image carousel, name, address, Open in Maps, description, trend score and rating, social posts (TikTok/IG links), and reviews.

Challenges We Ran Into

Making the API fast: No auth; optimized PostGIS queries and spatial indexes for nearby.
AI location extraction: Nemotron can miss or be unclear; added fallback Google Places text search and logging.
Natural language search: LLM-enhanced search is independent; falls back to text search if LLM is down.
Scoring across platforms: Normalized TikTok vs. Instagram engagement patterns separately.
iOS map interactions: Bottom sheets + filters while keeping the map interactive required clear state and cancellable calls.
Getting images: TikTok thumbnails can be poor and IG URLs expire; app tolerates empty images and uses placeholders.
Description generation: Switched Nemotron models when one was unavailable; generation is non-blocking.