QT - AI | Devpost

Inspiration

In an era where every keystroke is routed through centralized clouds, teams face a growing dilemma: they need the productivity boost of Generative AI, but they cannot risk data leaks, privacy breaches, or loss of intellectual property.

We asked a simple but radical question: Can we build a workspace as intelligent as modern SaaS tools, yet as private as a conversation in a soundproof room?

QT AI was born from the idea of Local-First Intelligence—where collaboration happens directly between peers, and AI acts as a private consultant that only sees what users explicitly choose to share.

What it does

QT AI is a fully browser-based, peer-to-peer collaboration platform powered by local-first architecture and Google Gemini 3.

Secure P2P Chat – Real-time messaging and file sharing over WebRTC with no central server.
Document Analyst – Upload PDFs or images and extract structured data like tables, entities, and summaries. Chat with your documents.
Second Brain (Decision Mode) – A reasoning engine for complex decision-making, analysis, and image understanding.
Idea Framework – A collaborative whiteboard for mapping ideas, systems, and workflows in real time.
Sheet Intelligence – An AI-powered spreadsheet editor that generates formulas, analyzes trends, and refactors data using natural language.
Outcome Workspaces – Structured templates (Decision Memos, Proposals) where AI acts as a consultant to challenge assumptions and draft content.

How we built it

QT AI follows a serverless, local-first architecture:

Frontend: React 19 + TypeScript for performance and type safety, styled with Tailwind CSS and custom neural animations.
Networking: WebRTC wrapped with PeerJS. Users connect via unique Peer IDs, and all synchronization happens directly between peers.
AI Integration: Google Gemini 3 (gemini-3-flash-preview) via @google/genai, using structured outputs (responseSchema) to generate UI-ready JSON instead of raw text.
Persistence: All data is stored locally in the browser using localStorage. No backend database is used.

Challenges we ran into

State synchronization without servers – Keeping chat, whiteboards, and live actions in sync required a custom peer-to-peer broadcast system.
Structured AI reliability – Ensuring consistent, valid JSON from the AI was challenging and required strict schema enforcement.
Large file handling – Sending PDFs and images over WebRTC data channels required careful optimization and local preprocessing.

Accomplishments that we're proud of

Zero-database architecture with $0 backend hosting cost and zero user data stored on servers.
The Neural Orb UI, giving the AI a responsive, living presence instead of a simple loading indicator.
Seamless multimodality, allowing text, images, PDFs, and spreadsheets to work together in a single flow.

What we learned

Local-first architecture builds trust – Users are more comfortable sharing sensitive data when it stays on their device.
Gemini 3 is fast enough for real-time UX, enabling AI-powered interactions without breaking flow.
Prompt engineering is UI engineering – System prompts were just as important as frontend components.