Meridian

Collaborative Data Analysis Tan-times faster

Problem You're Solving

Non-technical teams spend hours on repetitive data work: uploading spreadsheets, writing queries, waiting for results, losing track of how insights were discovered. Traditional tools (Julius AI, ChatGPT) work but aren't built for real-time collaboration or reproducible analysis workflows.

How It Works

Meridian is a collaborative data analysis platform where teams:

• Upload CSV/SQL data → instantly get auto-discovered insights
• Ask natural language questions → AI agents show step-by-step reasoning
• Watch results stream live as charts update in real-time
• See teammates' queries and analyses happening simultaneously
• Replay any analysis to understand HOW insights were discovered
• Make charts with AI and see them update live-time as data is being operated on

Notable Features

Live-Time Collaboration

• Real-time reactive updates via Convex subscriptions
• Multiple users querying simultaneously, all results sync instantly
• See query history + reasoning steps as they execute

Streaming Agent Reasoning

• Ask a question → agent breaks it into steps
• Each step streams to UI: "Reading column X..." → "Computing statistics..." → "Found pattern Y"
• Judges see transparent AI reasoning, not black-box results

Live-Updating tables

• Query/Mutate the data → charts update instantly
• Not batch processing like competitors
• Powered by Convex reactivity + TanStack Start streaming

Query Reproducibility

• Every analysis tracked with reasoning preserved
• Rollback system to previous queries
• Understand WHAT happened and WHY it happened
• Coming Soon: Git-like branching of data and merging

Vectorized OLAP Analytics

• DuckDB powers fast analytical queries
• Columnar storage optimized for aggregations
• Can analyze millions of rows instantly

Why I Built This

The problem: teams waste time on data instead of insights. The opportunity: combine real-time collaboration (Convex), streaming architecture (TanStack Start), and analytical power (DuckDB) to make data accessible without losing reproducibility.

Tech Stack

Frontend: TanStack Start + Mantine UI -- Helped me use DuckDB NODE instead of WASM, which made the app much faster and easier to use; Also used Tanstack Query and Table Backend: Convex (real-time, type-safe)
Data Storage: DuckDB (server-side via TanStack), Cloudflare R2 (file storage)
AI Agents: Gemini + Convex Agent Component with streaming steps
Data Integration: Firecrawl (URL → CSV)
Billing: Autumn (usage-based pricing)
Monitoring: Sentry error tracking + CodeRabbit reviews before commits Deployment: Netlify

Challenges We Ran Into

• Data Serialization Across RPC Calls
  • Issue: TanStack Start Server Functions needed to pass large datasets + DuckDB instances
  • Solution: Store data in Cloudflare R2/MotherDuck file storage, pass only references over RPC
  • Why it matters: Enabled server-side DuckDB (fast) instead of wasm (slow)
• Live-Time Updates at Scale
  • Issue: Firebase can't do live-time chart updates efficiently
  • Solution: Convex subscriptions handle reactive data flow + automatic cache invalidation
  • Result: Charts update milliseconds after query completes
• Server-Side DuckDB on Netlify
  • Issue: Netlify read-only filesystem + home directory requirements
  • Solution: TanStack Start's server/client separation allowed server-side DuckDB node
  • Why: DuckDB wasm is too slow; needed native performance

Why Meridian is unique

vs Julius AI: We have query reproducibility + streaming reasoning steps + live collaboration vs ChatGPT: Persistent analysis workflows + real-time team collaboration + transparent reasoning vs Traditional BI Tools: Natural language queries + AI reasoning + zero setup

Shipped Nov 17, 2025.