ROXBOX is hosted by AWS Amplify. Built with Next.js, the platform consists of a React frontend coupled with an ephemeral API route layer. The API routes serve as the central hub, interfacing with various external services to deliver the application's full suite of features. The application uses a multi-tier architecture and leverages a Cloud-to-Home Bridge Tailscale mesh network to support YouTube audio extraction and USB device synchronization.

ROXBOX was built to overcome the clunky experience found in all music players. Navigating the library is intuitive, offering seamless scrolling and filtering options by artist, album, or track. Users can effortlessly build playlists with a single click, which are then automatically synced to a USB device to facilitate offline car playback.

By long-pressing a track at the bottom of the screen, users can trigger the application to identify 10 acoustically similar songs and integrate them into a playlist. This feature offers the flexibility to either generate a completely new 10-song selection or append the discoveries to an existing playlist.

How it Works

The musiCNN (25 MB) deep learning model runs within an automated AWS Lambda function. When a track is uploaded to ROXBOX, this function extracts the audio's acoustic DNA, converts its unique sound profile into a 768-dimensional vector, and stores it in a Pinecone vector database. Because the model weights and heavy audio extraction dependencies exceed the standard 250 MB size limit of a Lambda function, the Lambda function is deployed as a Docker container.

When you long-press a song to find similar tracks, the Next.js API performs a real-time vector similarity search against Pinecone, returning matching results based on acoustic characteristics rather than text metadata. To ensure fresh discovery, the search engine automatically filters out other tracks by the same artist.

The application allows users to preview music from YouTube on a per-song or per-album basis. When a user decides to download a song, the audio from the song is extracted and automatically added to the user's S3 cloud collection. In album mode, you can easily select from a list of albums to preview and download music directly into S3. All newly added songs are synced immediately to the user's USB drive. The green dot indicator after the YouTube text in the title indicates that the home server is online and the USB drive is connected to the home computer. The indicator will turn yellow if the USB drive is not connected. The indicator will turn red if the home server cannot be reached.

The Audio Extraction Pipeline follows five sequential steps:

1. The Handoff: A payload containing the video URL, search query, original video title, and a unique jobId is sent from the Next.js frontend directly to the AWS Lambda Proxy function.

2. The Secure Tunnel: The Lambda Proxy spins up, securely joins a private Tailscale network, and acts as a lightweight proxy passing the JSON payload down to your home server.

{
  "url": "https://www.youtube.com/watch?v=R23D7Gndp7I",
  "action": "download",
  "title": "Oasis - Don't Look Back In Anger",
  "search_query": "oasis dont look back",
  "jobId": "extract_1777854482198"
}

3. Extraction: The Home Server reads the payload and invokes yt-dlp and FFmpeg to download the video stream and convert it into a high-fidelity MP3 file. A home server was required to get around the IP restrictions imposed by Google and the yt-dlp library. Since the audio extraction runs from a residential IP, this simply appears to look like somebody watching a video.

4. Agentic AI Processing & Metadata Cleaning: A local Llama 3.2 (2 GB) model reads the video context and original search intent. Governed by the Librarian Constitution (one page markdown file), the AI operates in an agentic fashion—reasoning about when and how to deploy its three specialized senses (AcoustID for audio fingerprinting, MusicBrainz for canonical releases, and YouTube Metadata for uploader context). This reasoning loop allows it to accurately resolve canonical metadata while filtering out noise and enforcing strict, USB-safe naming conventions.

5. Upload & Sync: The canonical audio file is uploaded to AWS S3 and automatically queued for a debounced synchronization directly to your USB drive, creating a nicely structured music library in the cloud and on a USB device.

Beyond traditional playback, ROXBOX serves as a dedicated performance training tool with a built-in implementation of the Norwegian 4x4 High-Intensity Interval Training (HIIT) protocol.

To optimize your training, the application natively manages workout timers for Warm-Up, 4-minute Burn intervals, and 3-minute Active Recovery sets. By synchronizing these intervals with specific playlists, ROXBOX ensures your highest-energy tracks hit exactly when your heart rate needs to peak.

The system provides a seamless experience where songs automatically fade out at the end of each interval. Users can fully customize these intervals in settings.

Gym-Proof Resilience

To ensure zero data loss during high-intensity sessions where cellular signals might be weak, the application implements an Offline-First Logging Queue. If a workout completes while the device is offline, the app serializes the workout data to localStorage. A visibility-aware Background Heartbeat monitors the connection status, automatically "draining" the queue and syncing logs to Google Sheets the moment the user returns to Wi-Fi or cellular range.

Upon completing a session, the app logs the date and interval lengths to a Google Sheet. A nightly automated process then queries the Whoop API to retrieve and store heart rate and performance data directly into the spreadsheet for comprehensive tracking.

To capture the nostalgic, tactile feel of physical media, ROXBOX provides two distinct user interface profiles designed for album enthusiasts: Classic Mode and Album Mode.

• Classic Mode: Presents a clean, information-rich interface optimized for quick queue management, listing tracks with high visibility.
• Album Mode: Reimagines the visual footprint by focusing heavily on the album artwork. Inspired by the 80s CD player experience, this layout makes the cover art the main event, filling the screen with bold, glassmorphic design and keeping the artist's visual intent front and center.

Building ROXBOX was an amazing journey. This entire product was created in three weeks using Google Antigravity, allowing me to quickly build and learn concepts that were completely outside of my wheelhouse.

On the downside, Antigravity was not perfect and would occasionally break things elsewhere in the application. But on the upside it would proactively do things for the better. For example, while styling a clean download button, it automatically unified and matched the style of all buttons across the application on its own accord.

Antigravity was highly productive. The similar song feature was implemented in an older application. I simply loaded both applications into Antigravity and instructed it to copy the similar song search feature. It completed this task instantly.

On the infrastructure side, UDP holepunching is a clever way to create a secure UDP mesh network bypassing network constraints enforced by home routers. Since every hostname on a Tailscale network must be unique, adding a small GUID to the hostname for the Lambda function ensures that multiple Lambda invocations will execute properly. It is also necessary to configure the Lambda functions as ephemeral hosts in Tailscale.

The agentic AI tuning process was cumbersome. Imagine tuning a guitar with a Floyd Rose tremolo for two hours! Looking back, a utility webpage should have been created to easily test and establish a ground truth with the AI.

Finally, Ollama proved to be an exceptional tool, treating AI models like Docker containers. Ollama is for Application Developers who just want to use a model instantly and reliably without any of the low-level math or driver overhead found in TensorFlow or PyTorch. Ollama makes it easy to load and run AI models using standard, stateless JSON web calls. Ollama llama. Really?