Founder & Lead Engineer: Yagnik Vanodiya
Status: Active Prototype / Backend Stable
Void AI is not a generic life-logger. It is a stealth, high-performance edge-computing ecosystem engineered specifically for academia.
Modern university lectures are dense, complex, and fast-paced. Existing AI recorders are optimized for corporate meeting action items. Void AI solves the academic bottleneck by acting as a continuous real-time lecture parser. It streams audio via Bluetooth Low Energy (BLE) to a custom-built processing pipeline, leveraging multimodal LLMs to dynamically generate structured study guides, extract core formulas, and build a searchable vector-database of a student's entire semester.
Note: This project builds upon the robust open-source Omi architecture, but the core infrastructure has been completely decoupled, overridden, and re-engineered with a proprietary backend and custom data schemas to support our academic focus.
Void AI operates on a distributed microservices architecture, spanning from bare-metal hardware to cloud-based AI generation.
1. The Hardware (The Edge)
- Current Iteration: ESP32-S3 with I2S microphones transmitting continuous audio chunks via BLE protocols.
- Next Phase: Custom-routed, miniaturized printed circuit boards (PCBs) designed in KiCad for optimal power efficiency and form factor.
- Future Expansion: Multi-sensor spatial analytics integration (e.g., BME680 environmental sensors and mmWave radar) for complete room occupancy and contextual environment tracking.
2. The Client (The Interface)
- Framework: Flutter / Dart (Cross-platform)
- State Management: Optimized for high-frequency WebSocket streams to prevent Out-Of-Memory (OOM) compiler crashes during continuous lecture recording.
3. The Backend (The Brain)
- Server: Python FastAPI deployed on a native Linux environment, securely tunneled via Ngrok for rapid local development.
- Speech-to-Text: Real-time streaming integration with Deepgram.
- AI Agent Pipeline: OpenAI integrations orchestrated via LangGraph for contextual memory extraction and RAG (Retrieval-Augmented Generation).
4. The Database (The Memory)
- Primary Store: Firebase Firestore with manually built composite indexes for rapid, complex query filtering.
- Caching layer: Upstash TCP database with SSL encryption enabled (replacing default Redis configs to prevent silent 500 internal server errors).
- Vector Search: Pinecone database for semantic retrieval of lecture concepts.
- Successfully bypassed upstream Linux C++ Out-Of-Memory compilation errors for the Flutter client.
- Decoupled upstream cloud dependencies, rerouting authentication and database logic to a proprietary Google Cloud / Firebase instance.
- Resolved hidden streaming errors by configuring custom Firestore composite indexes for AI memory retrieval.
- Stabilized the complete RAG text pipeline: Flutter → FastAPI → Upstash → Pinecone → OpenAI.
To prevent merge conflicts with upstream open-source branches, Void AI employs a strict Isolation Strategy. The main branch acts as a clean upstream mirror, while all proprietary logic, UI overhauls, and backend overrides are engineered and maintained in void-ai-main and dedicated feature branches.
Building the ultimate cheat code for the modern classroom.