Case Study

dataGenie — Multi-Agent Data Analytics Platform

dataGenie · dataGenie

MVP complete · local

Upload a CSV, ask in plain English, get back a chart, a written answer, and the SQL you can verify.

Overview

Project framing

A platform for the question 'what does this data say?' that doesn't require the asker to know SQL or even the column names. Upload a file, ask in plain English, get an interactive chart with a written answer and the underlying query you can verify. The agents handle the work, the user reads the answer.

Problem

Every non-technical teammate who needs data from a CSV either asks an engineer or fights Excel pivot tables. The engineering loop is slow; the Excel loop is error-prone and gets abandoned. The non-technical user ends up making decisions on intuition instead of data — not because they want to, but because the tooling demanded too much to use.

Role

Designer, engineer, evaluator. Solo build.

Why now

Combine those, and a real user can go from 'I have a file' to 'I have an answer' without ever seeing a query.

Topline

Constraints

Shipped as a single 154-file monolithic commit — feature-complete MVP, zero production hours.
Multi-provider fallback only earns its keep if the simple path stays cheap; the router has to be honest about complexity.

Outcomes

A non-technical user can go from 'I have a file' to 'I have an answer' without ever writing or seeing SQL — unless they want to check it.
Every answer comes with the underlying query and a verification pass, so results are checkable rather than taken on faith.
Follow-up questions work in context — the system remembers the current dataset and prior results, so 'now compare those regions' just works.
Feature-complete MVP running locally as a full stack, wired and tested end to end.

Architecture

Step 1

Ingest

Upload CSV / Excel / PDF / connect a database.

Step 2

Route

The complexity router classifies the question as simple, moderate, or complex — the routing decision is what makes the cost curve work.

Step 3

Execute (simple)

No agents, no tokens spent.

Step 4

Execute (complex)

Orchestrator coordinates.

Step 5

Verify

Verification agent cross-checks the others' work. Hallucinations get caught before they surface.

Step 6

Answer

Interactive Plotly chart + written explanation + confidence score + the SQL you can verify. Export to PNG / SVG / CSV / Excel / PDF.

Multi-turn conversation

1,193-line engine tracks current dataset, columns mentioned, and prior results so 'compare those regions' actually works.

OutcomeOne question, one chart, one written answer, one verifiable query.

Key Decisions

A dedicated Verification agent

Add a fifth agent whose only job is to cross-check the other four.

Multi-agent systems that don't verify their own outputs hallucinate confidently. Catching that has to be part of the system, not a hope.

Extra tokens on every complex query. Worth it because the alternative is silently wrong answers.

Struggles

Scope. Shipped as a single 154-file monolithic commit with 448 tests passing on my laptop and zero production hours. Feature-complete MVP with no real users.

No clean fix — this is a habit problem, not a bug. The recovery is visible in how I'm now shipping duSraBheja: smaller PRs, earlier deploys, letting real usage drive the next round of agents rather than designing them up front.

Learnings

Multi-agent systems need a verification layer before they need a sixth agent.

Hybrid routing is a product decision, not a technical one — it's about when to spend tokens.

Feature-complete-but-unreleased is a worse state than partially-shipped.