Get Magic
Get Magic — Magic Assistant & Open-Floor Assistant
AI co-pilot for getmagic.com and embedded assistant surfaces — enterprise extension workflows with RAG, memory, and real tool actions.
Timeline
2023 – 2024
Stack size
8 technologies
Role
Senior AI Engineer (Consulting)
Company
Get Magic (getmagic.com)
Overview
Overview
Magic is an executive-assistant service combining vetted humans with AI tooling. The "Magic Assistant" is an AI co-pilot that drafts, summarises, and executes for the human assistant; the open-floor assistant is a team-wide helper any assistant can call on.
- I built features across both — the Plasmo Chrome extension surface, in-browser RAG over playbooks and prior tickets, and MCP-style tool servers wiring agents to CRM, calendar, email, and internal Magic APIs.
- I also contributed to in-product client surfaces: dynamic sidepanel tooling, internal feature-flag controls, and competitive-intelligence views that kept assistant workflows discoverable and fast.
Problem
Problem
Executive assistants live in dozens of tools — Gmail, Calendar, CRMs, client docs. Tab-switching to a separate AI chat breaks flow. Magic needed AI to live where the work happens with full context and the ability to take action.
Architecture
Architecture
Extension
AI
Tooling
Infra
Features
Features
- Context-aware Magic Assistant in any web page.
- Open-floor team-wide assistant for cross-assistant knowledge sharing.
- In-browser RAG with IndexedDB so sensitive context stays on-device when possible.
- MCP-style tool servers for CRM, calendar, email, and internal Magic APIs.
- Edge-deployed compute on Cloudflare Workers and AWS Lambda for low latency.
- Enterprise sidepanel integrations with dynamic tool loading, internal settings gates, and prompt-aware tool routing.
- Competitive-intelligence dashboard surfaces combining brand, social, and report views for assistant research workflows.
AI systems
AI systems
- Hybrid retrieval over playbooks, tickets, and client preferences.
- Agents with scoped tool access and full audit trails.
- Privacy-first design — data minimisation, PII scrubbing, and least-privilege tool calls.
- Prompt-safety processor layers (detection + scrubbing) to reduce prompt-injection risk in tool-enabled assistants.
Outcomes
Outcomes
- Reduced research and drafting time per assistant ticket.
- Standardised reuse of playbooks across the assistant team.
- Improved reliability of multi-tool assistant runs through scoped gateway patterns and feature-flagged rollout.
Technologies
Technologies
FAQ
FAQ
Why an extension instead of a web app?
Assistants already work in Gmail, Calendar, CRMs, and client tooling. The extension gives the AI page context, selection context, and direct tool actions without breaking the user’s flow.
More work