Quiver | Devpost

Quiver1
Quiver2
Quiver3
Quiver4
Quiver5
Quiver6
Quiver7
Quiver8

Inspiration

As a freelance software developer juggling work and a computer science degree at an open university, I struggled to balance work and studies. Unstructured materials, large PDFs, and poor internet connectivity while commuting around my local area, and sometimes at home, along with expensive data. In addition, many AI tools for studies/summaries were not really built for studies, especially after having a long chat, often crashing, or requiring constant internet. I saw the same problem faced by students, busy professionals, and parents earning degrees. Quiver was born to solve this: an offline-first study app.

What I Learned

Hybrid AI Approach: All features except content (summaries, quizzes, and flashcards) generation currently work offline (future versions will gradually be optimized to work fully offline, including the content generation by the AI).
On-Device AI Testing: I am testing on-device AI models locally to handle large contexts (large numbers of words). The goal is to make the entire process offline while keeping performance smooth.
Voice Technology Research: Exploring text-to-speech and speech-to-text solutions such as Vosk, Mozilla DeepSpeech.
Model Fine-Tuning: Learning techniques like LoRA (Low-Rank Adaptation) to fine-tune models for efficient on-device inference.
Cross-Platform Porting: Researching how to integrate AI models into Flutter for Android, iOS, and upcoming desktop versions using C++.
Exploring Liquid AI (LEAP & Apollo): These initiatives bring powerful, lightweight AI models and developer tools for edge devices. Their focus on speed, efficiency, and multi-platform support aligns perfectly with Quiver’s roadmap for offline AI, and it's also open source.