Structured-output ingestion from unstructured social content

Media, Audio & Communications

Industry: Media, Audio & Communications

Platform: iOS + Node.js backend

Stack: Swift · SwiftUI · Supabase · Node.js · LLM API

Description: Native iOS client over a large PostgreSQL set with AI structured extraction from social URLs.

---

The engineering surface combined data-scale performance work on a 138K-record PostgreSQL set against a mobile client with a server-side pipeline that extracts structured records from unstructured social-media content. The Explore screen required moving the perceived load off the network entirely — a curated in-binary index serves the first render synchronously, while a background task group enriches sections from the server. The ingestion pipeline shells out to a metadata extractor as a child process, routes results through an LLM with a strict structured-output prompt, and detects ephemeral CDN URLs to re-upload to permanent storage before persistence.

We owned the iOS client, the Node.js extraction service, the PostgreSQL schema with GIN-indexed arrays and composite sort indexes, and the LLM prompt design. The client owned the product concept and the recipe data set.

Engagement: ~4-month build · 2-engineer team (iOS + backend) · 2025.

Engineering signal: Explore screen 15-20s → sub-100ms cold start via in-binary index; parallel withTaskGroup image uploads cut multi-photo save 25s → ~5s.