iOS / AI

Relora

An AI that reads between the lines of love texts. OCR runs on the device, the psychological analysis runs on a cloud LLM.

SwiftSwiftUIAWS BedrockAWS CDKSwiftData

Open the app Related articles →

Overview

Relora reads a single screenshot of a chat thread, reconstructs the conversation, and surfaces the emotional temperature and possible next moves. OCR is fully on-device via Apple Vision; analysis and reply suggestions run on Amazon Bedrock with Qwen3 or Claude Sonnet 4.6.

The problem

Stepping back from a chat with someone you care about is hard, and there are not many people you can ask. At the same time, pasting a raw chat history into a cloud AI feels deeply uncomfortable.

Key features

Reconstructs the conversation from a single screenshot
On-device OCR via Apple Vision keeps the raw text local
Surfaces relationship temperature and risk points in one view
Generates multiple reply candidates tailored to the situation
Free / Pro tiers powered by StoreKit 2

Video & channels

Relora's short videos compress where the app fits and what its relationship analysis outputs into 30 seconds, with captions and titles added in CapCut. The same source ships to YouTube Shorts, TikTok, and Instagram Reels — an entry point that introduces the product to site visitors.

Same video also on TikTok Instagram .

Tech stack

iOS

Swift, SwiftUI, SwiftData, Apple Vision

Backend

API Gateway, Lambda (Python), Cognito, DynamoDB

LLM

Amazon Bedrock — Qwen3 Next 80B, Claude Sonnet 4.6

IaC

AWS CDK (TypeScript)

Architecture

The iOS app uses SwiftUI with MVVM and SwiftData for persistence. Images are OCR-processed on-device with Apple Vision, and only the text is sent through API Gateway + Lambda to Amazon Bedrock. Infrastructure is defined in AWS CDK (TypeScript); Cognito gates access and Bedrock Guardrails contain prompt safety.

AI's role and boundaries

AI handles the psychological analysis itself, routing between Qwen3 Next 80B (free tier) and Claude Sonnet 4.6 (paid tier) on Bedrock by role. A psychology-grounded system prompt and per-language guidance hold analysis quality steady, and prompt templates ship as a Lambda Layer. Remove inference and no core feature remains, so the product is built on AI by design. Reading screenshots stays on-device OCR, and the model is never asked to identify a person or make judgements about a relationship. Quality is checked by watching Bedrock latency and by tuning prompts and Guardrails based on reviews and support requests.

Evaluation & Operations

Since launch we've watched Cognito request patterns and Bedrock latency. Prompt and Guardrails tuning is driven by App Store reviews and support requests.