S.N.A.P. (Semantic Notes & Action Processor)

Inspiration: Every day, millions of meetings produce hours of unstructured conversations—but only a fraction of decisions, tasks, and deadlines are actually captured. We noticed that teams often leave meetings without clear ownership, forgotten action items pile up, and follow-ups fall through the cracks. The gap between talking about work and doing the work is massive.

The inspiration for S.N.A.P. (Semantic Notes & Action Processor) was to eliminate "meeting amnesia" entirely. We wanted to build a tool that not only summarizes discussions but automatically turns raw audio into a crystal-clear, actionable workflow.

What it does S.N.A.P. is a local-first, AI-powered meeting assistant. You paste (or dictate) your raw meeting transcript, and our engine instantly extracts:

Structured action items with inferred owners, deadlines, and priorities. Key decisions recognized from the conversation. A Meeting health score that measures the clarity, actionability, and participation balance. A split-screen Meeting Comparison tool to track if your team is resolving action items week-over-week. Crucially, S.N.A.P. pushes these insights directly into your workflow via our Webhook Automation Engine (perfect for seamless integration with n8n and Relay.app). With single-click exports, you can push tasks to Google Calendar, format markdown blocks for Slack, export Notion-ready lists, or beam JSON to any pipeline.

How we built it We built the frontend as a highly responsive single-page application using React 18 and Vite.

UI & Design: To give it a premium feel, we created a custom "Aetheric Void" design system featuring glassmorphism, grain textures, and interactive, living metallic blobs using Three.js and React Three Fiber. AI Engine: The core semantic processing is powered by the Google Gemini 2.5 Flash API. We engineered complex, robust system prompts so the LLM reliably returns structured data out of unstructured conversations. Voice Integration: We utilized the browser's native Web Speech API for real-time dictation with visual feedback. Deployment & Privacy: The entire application is deployed globally on Google Cloud App Engine. It operates entirely local-first—all analysis and history are kept securely in the browser's local Storage.

Challenges we ran into One significant hurdle was getting an LLM to consistently output perfectly structured JSON data with accurately inferred context (like mapping "I'll do that before the weekend" to an exact deadline and assignee). We spent hours refining the prompt strategy for Gemini to guarantee zero dropped tasks.

Another challenge was integrating the Three.js canvas so that the metallic background effects looked stunning without causing UI latency or crashing the browser native speech-to-text listener.

Accomplishments that we're proud of We're exceptionally proud of the Webhook Automation Engine. Allowing a user to seamlessly connect one webhook URL and immediately start piping structured meeting intelligence to n8n or Relay.app creates endless possibilities.

We are also really proud of the local-first "Meeting Comparison" diff tool—creating an intelligent system that accurately realizes which topics from last week carried over into this week, all without needing a traditional backend database!

What we learned We vastly deepened our understanding of modern prompt engineering and how capable lightweight, flash LLM models (like Gemini 2.5 Flash) can be for complex data extraction. We also learned advanced techniques for blending HTML/CSS UI layers perfectly over an interactive WebGL canvas.

What's next for S.N.A.P. We plan to introduce deeper native integrations (like full Google Workspace OAuth rather than relying entirely on URL generation and webhooks), adding direct PDF/Audio file uploads, and creating real-time multiplayer meeting spaces so teams can co-edit notes before the final export.