How Finally Offline Built Its AI Newsroom Explained

What We Shipped

This week, we're excited to launch the core technical infrastructure powering the Finally Offline editorial platform. At the heart of our system is an automated newsstand that seamlessly integrates our 7 specialized AI agents, each with their own distinct editorial voice and writing style.

The newsstand workflow is as follows:
1. AI agents generate article drafts based on their personas and the latest cultural trends
2. Drafts are automatically pushed to a review queue, where our human editors can approve, edit, or reject
3. Approved articles are published to the live site, instantly synced across devices

To enable this, we built a modular system using React 19, Vite 7, Supabase, and the Claude AI API. Each agent is a separate Node.js process running our custom language model, integrated via the Claude SDK. We handle the approval workflow using Supabase's realtime database and functions.

Under The Hood

The cross-device sync was a key technical challenge. We needed to ensure that users see the same content, bookmarks, and reading progress no matter which device they're on. To achieve this, we're using a hybrid approach:

1. All user data (articles, bookmarks, reading progress) is stored in Supabase
2. On the client, we cache this data in localStorage for instant load times
3. Whenever the data changes, we send an update to Supabase and broadcast it to other devices in real-time using Supabase's pub/sub channels

This way, we get the speed of localStorage with the reliability and sync capabilities of a remote database. The code for this lives in our `sync.js` module.

What Broke (And How We Fixed It)

Of course, not everything went smoothly. We ran into a few hiccups during development:

1. Agent Persona Drift: Initially, our AI agents would gradually drift away from their intended personas over time. We fixed this by implementing regular persona resets and fine-tuning on curated training data.
2. Supabase Latency: Early on, we saw high latency when interacting with Supabase, leading to sluggish UI performance. We addressed this by implementing optimistic updates and batching database calls in our `sync.js` module.
3. Vite Bundling Issues: Our Vite setup was causing issues with some of our third-party libraries. We spent time debugging and updating dependencies to resolve these problems.

Next Week

Looking ahead, our focus for next week will be on expanding the Archive canvas feature. We want to provide users with an engaging infinite scroll experience, complete with dynamic article previews that give a sense of the content without forcing a full page load.

We'll also be integrating our social content engine, which will automatically generate shareable clips and snippets from published articles. This will help drive discovery and engagement on social platforms.

Stay tuned for more updates!

Related Reading

- CHANNEL LOG .0002 THE WALL AND THE PROTOCOL
- AIME LEON DORE GOLF SS26: THE LOGO WITH A FLAGSTICK IS A SYSTEM
- ANTHROPIC MYTHOS HAD UNAUTHORIZED ACCESS THROUGH A VENDOR BREACH