01CASE STUDY

AI System Design

GHOST-AI

ROLE

SOLO BUILD — DESIGN TO PRODUCTION

TIMELINE

SHIPPED · ACTIVELY MAINTAINED

SCOPE

Realtime architectureAgent tool-callingCanvas engineSpec generation

STACK

Next.jsTypeScriptLiveblocksGeminiTrigger.dev
GHOST-AI — screenshot

OVERVIEW

Describe a system in natural language and watch an autonomous AI architect draw it — nodes, edges, labels — on a canvas your whole team is editing live.

Ghost-AI is an enterprise-grade collaborative whiteboard built on CRDTs: every shape, edge and label lives in Liveblocks LiveMaps, so simultaneous edits from multiple collaborators merge without conflicts. Live cursors, presence avatars and an in-workspace chat feed make it feel like one shared room.

The AI architect is a Gemini 2.5 Flash agent wielding eight discrete tools — addNode, moveNode, resizeNode, updateNodeData, deleteNode, addEdge, deleteEdge, finalizeDesign — executed serverlessly through Trigger.dev so long-running generations never hit timeouts. Canvas state auto-saves on a two-second debounce to Vercel Blob, indexed in PostgreSQL via Prisma.

Finished designs compile into multi-page Markdown specifications — overview, architecture, data flow, technology choices — turning a sketch session into technical documentation automatically. Auth and multi-tenant workspaces run on Clerk with room-level isolation.

OUTCOMES

8agent tools driving the canvas
CRDTconflict-free multiplayer editing
17releases shipped
The diagram isn't documentation of the design. The diagram is the design.
© 2026 RUSHIKESH PALANDEBACK TO TOP ↑
© 2026 RUSHIKESH PALANDE. ALL RIGHTS RESERVED.