Skip to content
Lubo Bali
All projects

AI agentsAny business

mergeAI

Upload spreadsheets, ask a question in plain English, get a checked answer

~8 sfrom question to a checked answer

Try the live demo (opens in a new tab)Watch the demo (opens in a new tab)See the code on GitHub (opens in a new tab)

Proof you can check:Live demoCode on GitHubVideo walkthrough

AI agents

Ask in plain English, a checked answer in ~8 s

~8 squestion to checked answer
  1. 1Schema agentfinds how your files connect: EmpID is Employee ID
  2. 2SQL agentwrites a real PostgreSQL query to join them
  3. 3Validatorruns it and checks the result, with no AI in it
  4. Zero rows or a wrong column? Back to the agents, up to 3 rounds.

Lubo Bali · lubobali.com

No. 013 agents that check each other

The problem

Joining two spreadsheets that were never made to match takes drag-and-drop, formulas or SQL. Most people just want to ask the question and trust the answer.

What I built

Three AI agents: one finds how the files connect (EmpID is Employee ID), one writes a real PostgreSQL query to join them, and a checker runs it. Zero rows or a wrong column? It sends feedback and they retry, up to 3 rounds.

Tech & why

Real SQL in a real database instead of an AI guessing in memory, so every answer can be checked: "View SQL" shows exactly what ran.

  • Next.js
  • TypeScript
  • PostgreSQL
  • NVIDIA Nemotron
  • Vercel

The result

Any CSV becomes something you can ask questions in plain English, with a chart, a table and a short summary in about 8 seconds. It's live: try it with your own files.

See the real app

mergeAI: three agent cards (Schema, SQL, Validator) all done, the question, a short summary and a training outcome pie chart
Real mergeAI screen, demo HR files: the three agents done, then the answer.

How this agent is built

Every AI agent has the same five parts. Here is what each one is in mergeAI, so you can judge any agent the same way.

  1. ModelNVIDIA Nemotron Ultra 253B writes the SQL; Nano 8B reads the files and writes the summary.
  2. HarnessThree agents in a loop (schema, SQL, validator), shown live, up to 3 rounds of self-correction.
  3. GuardrailsA validator with no AI in it runs every query and sends back zero rows, wrong columns or case mismatches before you see an answer.
  4. KnowledgeYour CSV files in PostgreSQL (one table for any shape of file) plus a map of how the files connect.
  5. ToolsReal PostgreSQL queries (joins, CTEs), charts, and Excel or CSV download.

Similar projects

Have a problem like this? Get your free AI blueprint in about a minute.

Tell me your problem. I'll find a way to build the solution.

Available now

Book a free 30-min call (opens Calendly in a new tab)

No obligation. Pick a time that suits you.

Prefer email?data@lubobali.com