Souvik Biswas has released Sloth Reads, a Hacktoberfest submission that tackles the universal developer problem of the unread bookmark pile. The tool leverages Gemma 4 to process saved articles overnight, generating summaries, takeaways, and verdicts so users can triage their reading lists without manual effort. Built with a focus on open innovation, the project demonstrates how small, scheduled inference tasks can deliver high-value utility at negligible cost.

The Overnight Triage Pipeline

The architecture relies on DigitalOcean Functions to wake the sloth every hour between 1 AM and 5 AM. During each window, the system processes a batch of articles using Gemma 4, specifically the gemma-4-31B-it model on serverless inference. For each item, the LLM outputs a JSON object containing a 2-3 sentence summary, three key takeaways, and a binary-style verdict: 'worth reading,' 'skim it,' or 'skip it.' Crucially, the verdict includes a personalized justification based on a one-line user profile, distinguishing it from generic summarization tools. The system uses bge-m3 embeddings for semantic search, allowing users to query their pile by meaning rather than exact keywords. At 6 AM, a final function call aggregates the night's findings into a one-minute morning brief. The infrastructure is lean: App Platform hosts the containerized app, while Tiger Data (TimescaleDB) manages the state using hypertables for run logs and continuous aggregates for tracking time saved.

Cost and Performance Metrics

Biswas highlights the economic viability of open models for this use case. Running Gemma 4 on DigitalOcean costs $0.18 per million input tokens and $0.50 per million output tokens. Processing an average article costs approximately one-tenth of a cent, meaning an entire night's reading pile costs less than a single cent. This low barrier allows the model to read everything, including items the user will likely never engage with, maximizing the utility of the triage process. In a test against Biswas's own backlog of 19 articles from 2022 and later, the sloth correctly identified four outdated hackathon announcements as 'skip' items. The tool claimed to save 52 minutes of reading time by filtering out noise and providing concise summaries for the remaining items. The morning brief itself took only 1 minute and 11 seconds to listen to, providing a high signal-to-noise ratio for daily planning.

The Voice Integration Caveat

While the core reasoning engine is open, the user experience relies on a closed proprietary service: ElevenLabs. The 'Read & listen' feature uses ElevenLabs' streaming endpoint with character-level timestamps to highlight words in real-time. Biswas notes that while open text-to-speech options are improving, they currently lack the combination of natural voice quality and precise timing metadata required for this specific feature. The implementation streams audio in paragraph-sized chunks to minimize cost, stopping generation immediately when the user pauses.

Key Takeaways

  • Gemma 4 on serverless inference enables high-volume, low-cost document triage for personal knowledge management.
  • Hybrid search using bge-m3 embeddings and full-text indexing allows users to retrieve old bookmarks by semantic intent.
  • Open-source LLMs can replace manual reading for low-value backlog items, saving significant time with minimal financial overhead.
  • Proprietary APIs like ElevenLabs still hold an edge in niche UX features like synchronized word highlighting, despite open model progress.

The Bottom Line

Sloth Reads proves that you don't need massive context windows or agentic loops to solve personal productivity problems. By restricting the model to a narrow, high-frequency triage task, the developer achieved a practical, low-cost solution that actually respects the user's time. It is a refreshing counterpoint to the trend of over-engineering AI assistants for simple utility tasks.