Utkarsh G19, a student developer, has released ReadYou, a tool designed to demystify Hacktoberfest winning projects by generating technical briefings, audio summaries, and interactive chat sessions. Built for the Hacktoberfest Open-Source AI Challenge Week 3, the application addresses a common pain point for developers: the time and cognitive load required to understand unfamiliar open-source architectures. Instead of relying solely on hosted APIs, ReadYou runs Gemma 4 E2B locally via Ollama, offering a practical example of how student-level hardware can support meaningful AI experimentation.

The Pipeline: From Noise to Signal

The core innovation in ReadYou isn't just summarization; it's rigorous source validation. The initial version of the tool fed linked GitHub repositories directly into Gemma, which often resulted in hallucinations because the links pointed to dependencies or unrelated frameworks rather than the actual submission. To fix this, Utkarsh implemented a repository relevance check that treats linked READMEs as candidates rather than trusted sources. The system now uses Gemma to compare these candidates against the original DEV submission, requiring a confidence score of at least 0.85 before including any external documentation in the summary prompt. If the classification is uncertain, the tool defaults to the original submission text, preventing the model from describing features that don't exist.

Voice Mentor: Conversational Code Exploration

Beyond static summaries, ReadYou integrates ElevenLabs to create a 'Voice Mentor' feature. This pipeline allows users to ask questions about a project by speaking or typing, with the system transcribing the query, reasoning through the answer using local Gemma 4, and synthesizing a spoken response. The interaction is sequential rather than real-time speech-to-speech, but it includes state indicators for listening, thinking, and speaking. Users can interrupt spoken answers to ask follow-up questions, and the backend ensures that spoken responses are cleaned of Markdown headers and URLs for better audio clarity. This approach aims to make technical deep dives feel like informal conversations with a developer friend rather than reading a dry report.

Honest Limitations and Testing

Utkarsh is transparent about the current limitations of the tool. While the backend passed 45 automated tests, these primarily verify how the application handles model output and evidence citation, not the real-world accuracy of Gemma 4's reasoning. There is no systematic benchmark yet for the model's performance across a fixed set of projects and paraphrased questions. The developer notes that smaller models like Gemma 4 E2B can be repetitive or overly cautious, and latency remains a challenge due to the sequential nature of the voice pipeline. However, the focus on inspecting prompts and changing retrieval logic demonstrates a valuable lesson in building reliable AI applications: sometimes the fix isn't a better prompt, but changing what data the model is allowed to see.

Key Takeaways

  • ReadYou uses a 0.85 confidence threshold to filter irrelevant GitHub links before summarizing Hacktoberfest projects.
  • The tool runs Gemma 4 E2B locally via Ollama, reducing dependency on hosted APIs for reasoning tasks.
  • Voice Mentor combines ElevenLabs speech recognition and synthesis with local LLM inference for interactive project exploration.
  • Automated tests verify citation accuracy and scope assessment, but systematic live-model evaluation is still pending.

The Bottom Line

ReadYou proves that local LLMs can handle complex retrieval tasks without expensive cloud APIs, provided the developer builds robust validation layers to keep the model honest. It’s a refreshing look at student-led innovation that prioritizes understanding over hype.