NOVA AI
Website
A desktop automation companion that listens, understands, and acts. Control your system volume, brightness, and web interactions strictly through voice.An end-to-end AI framework designed to assist you in your daily tasks. We will be scale it by making this AI effiecient enough to code as well.
Client:
HackXtreme Hackathon
Role:
Product Designer
Year:
2025
Explore the full story –
Challenge
Cloud-based AI assistants suffer from high latency, privacy risks when handling voice and personal data, and are vulnerable to prompt injection attacks. Monolithic AI architectures lack modularity, making them difficult to scale and secure. Existing voice assistants rely on centralized cloud processing, which introduces network delays exceeding 500ms and exposes sensitive user data to third-party servers. Additionally, adversarial inputs can bypass naive security filters, allowing attackers to manipulate AI behavior through crafted prompts — a critical flaw in systems that lack a dedicated pre-inference security layer.
Objective
The objective of MODEL-X (NOVA) was to design and build a modular, edge-first AI assistant architecture that combines low-latency local inference with a zero-trust security gateway. Specifically, the project aimed to: (1) implement a real-time voice-to-voice pipeline with speech-to-text and text-to-speech capabilities, (2) introduce a pre-inference security layer that detects and blocks prompt injection attacks before they reach the LLM, (3) build an agentic orchestration engine that dynamically routes queries to specialized sub-models (reasoning, retrieval, tool execution), and (4) achieve sub-200ms end-to-end response latency while maintaining full user data privacy through on-device processing.
Results
NOVA successfully delivered a working end-to-end AI assistant pipeline comprising six stages: voice capture, speech-to-text, security gateway, orchestration, LLM processing, and text-to-speech. The security layer demonstrated effective detection and blocking of adversarial prompt injection attempts before inference. The modular orchestration architecture dynamically decomposed complex queries and routed them to specialized sub-models, with results aggregated for coherent responses. The system achieved a sub-200ms pipeline target through quantized models, asynchronous processing, response caching, and parallel module execution. Additional capabilities included live weather/news fetching without API keys, Codeforces problem solving with real statement fetching, background thread logging to avoid blocking the streaming response, and a desktop automation companion for volume, brightness, and web control via voice — all while operating under a zero-trust, privacy-first design principle.
