KanbanSync

Inspiration

Modern datacenter engineers operate at massive scales while still using an outdated workflow, regarding ticketing, task planning, routes, etc. To make this process more efficient, we decided to build an AI-powered system that will actually work with the engineers and not against them. This system understands natural language, validates tasks using real technical manuals, optimizes physical workflows, reduces downtime, and makes ticketing modern. We built KanbanSync to bring AI-assisted operational intelligence directly into the daily lives of data center technicians, making their work easier and more efficient.

What it does

KanbanSync is an intelligent, real-time workflow assistant for datacenter operations, combining a drag-and-drop Kanban board with a powerful AI co-pilot:

Converts natural-language instructions into structured service tickets
Validates procedures using a dual-RAG engine (Technical RAG + Datacenter RAG)
Suggests required inventory, components, and procedures automatically
Prioritizes tasks using an LLM with transparent reasoning
Generates optimized “work blocks” to minimize walking paths inside the datacenter
Provides a conversational sidebar for queries about inventory, procedures, or hardware
Updates the board in real-time using WebSockets

How we built it

Frontend

Next.js + React.js for the full interface
TailwindCSS for UI styling
Framer Motion to create smooth card animations and sidebar transitions
TypeScript for type safety and clean component design
WebSockets for live status updates on tickets

Backend

FastAPI for a clean, scalable API layer
Supabase for user authentication, database storage, and vector embeddings
OpenAI GPT-4o-mini to: -Parse natural language into structured ticket fields -Assign priorities with explanations -Synthesize RAG results into human-readable validation feedback
OpenAI Ada embeddings for vector search
Custom dual-RAG pipeline (Technical + Datacenter RAG) built with manual vector search and LLM synthesis
Supabase RPC functions for similarity matching
Pydantic models for robust data validation and serialization

Challenges we ran into

Non-Technical:
Misunderstood the original prompt and causing us to pivot from our original idea completely
Stay awake and finding a place to work efficiently

Technical:

backend complexity — wiring Supabase, FastAPI, RAG, and LLM routes together with clean imports
Vector search tuning — finding the right match thresholds and embedding models
Parsing natural language reliably — GPT needs a ton of guardrails to produce consistent formats
Real-time board updates — syncing frontend and backend WebSockets
Time constraints — building a full dual-RAG system + full UI on low time
Clerk / Supabase auth integration — webhook imports and package paths broke several times

Accomplishments that we're proud of

Built a functioning dual-RAG validation pipeline
Fully operational natural-language: structured ticket LLM system
Smooth, polished Kanban UI with drag-and-drop
Created real-time work block optimization using AI
Implemented a live conversational assistant integrated with procedures and inventory
Built a clean backend architecture