Vision and What the Agent Does
Most devotional apps put the work on you: open the app, pick something, read it. I wanted to flip that. Verse of the Day is an always-on agent that does the picking, the writing, and the art direction on its own, every single day, and hands you the finished thing.
Every evening at 8:45 PM, an Amazon EventBridge Scheduler rule fires and invokes a Lambda function with no manual trigger behind it. Nobody opens anything, nobody clicks anything. That function selects a Bible verse from a public-domain translation (KJV/WEB, specifically to avoid the copyright issues that come with NIV/ESV), checks it against a rolling history so the same verse doesn't repeat within a month or two, writes an original short reflection and a short prayer around it, generates a fresh piece of artwork to match the day's mood, and pushes a notification out to every subscribed device. By the time anyone checks their phone, the day's edition already exists: verse, reflection, prayer, and art, all generated fresh, none of it pulled from a template.
Opening the notification drops you into a three-slide carousel: the verse itself, the reflection, then the prayer, all sharing one visual mood for that day that's different from the day before.
How I Built It
The part I spent the most time thinking through before writing any code was the distinction between "scheduled" and "autonomous and creative." A cron job that rotates through a fixed set of pre-made verse graphics would technically run on a schedule, but it isn't actually making anything, it's just displaying. So the whole pipeline is built around real generation happening every cycle, not content selection from a static pool.
Amazon Bedrock does three jobs in one structured call: pick the verse, write the reflection, write the prayer, and also produce a short art-direction prompt for the day (palette, mood, style) rather than a literal description. Getting Bedrock to return one clean JSON object: verse, reference, reflection, prayer, image prompt, instead of parsing several separate free-text responses saved a lot of pain later. It meant one call, one parse, one place for things to go wrong instead of four.
AWS Services Used / Architecture Overview
Amazon EventBridge Scheduler fires the pipeline daily at 8:45 PM;
AWS Lambda, runs the full generation pipeline: calls Bedrock, calls Nova Canvas, writes to DynamoDB, sends the push notification;
Amazon Bedrock (Nova Lite/Micro): verse selection, reflection, prayer, and image-direction prompt, returned as one structured JSON object;
Amazon Nova Canvas generates that day's background art from the image-direction prompt;
Amazon DynamoDB stores each day's edition (verse, reflection, prayer, image reference) and all push subscription records;
Amazon S3 stores the generated images;
Amazon API Gateway + Lambda small read API the PWA calls to fetch the current (and past) editions;
CloudFront / Amplify hosts the PWA frontend
Flow: EventBridge triggers the Lambda → Bedrock generates the day's content and art direction → Nova Canvas generates the image → both get written to DynamoDB/S3 → the Lambda reads all stored subscriptions and sends a Web Push notification to each → the PWA fetches and renders today's edition as a three-slide carousel when opened.
What I Learned
The biggest thing I walked away with: AI coding agents have genuinely shifted what's possible in a short window. What used to take weeks: going from a rough idea to a working, deployed prototype, happened in a couple of days here. The bottleneck isn't writing code anymore; it's clarifying what you actually want built. That's a different kind of skill to sharpen going forward.
Loading article