How Would You Architect a Data Analysis Agent? Interview Guide

Interview Scenario

The interviewer asks how to design a data analysis agent, prompting a four‑module architecture.

Four Core Modules

1. Intent Recognition

This module parses the user's question to identify the analysis goal, such as trend analysis, comparative analysis, or anomaly detection, and determines relevant metrics, time ranges, and dimensions.

2. Data Acquisition (NL2SQL)

The core task is converting natural language into SQL (NL2SQL) and executing the query. It supports multiple data sources, including SQL databases, Excel files, and API endpoints. Generated SQL must undergo safety checks to prevent injection attacks.

3. Analysis Execution

Time‑Series Analysis : applies statistical methods or forecasting algorithms like Prophet.

Correlation Analysis : computes correlation coefficients.

Data Visualization : creates charts using Matplotlib or ECharts.

4. Result Presentation

Automatically selects appropriate charts based on data type.

Generates natural‑language summaries that explain findings rather than merely listing data.

Supports follow‑up questions, e.g., “Why did sales drop in March?”

Improving NL2SQL Accuracy

Schema Description Optimization : clearly document business meanings of table and column names.

Few‑Shot Examples : provide several correct NL‑SQL pairs as prompts.

Error‑Feedback Iteration : automatically correct SQL after an error response.

Manual Confirmation for Complex Queries : route high‑risk queries to human reviewers.

Real‑World Scenarios

E‑commerce Data Analysis : user asks for weekly sales trend → NL2SQL generates query → execution → trend chart with conclusion.

Operations Dashboard : user asks which channel has the highest conversion rate → NL2SQL → execution → comparative chart.

Financial Analysis : user asks why Q4 profit declined → NL2SQL → multi‑dimensional analysis → reason‑analysis report.

Interviewers often probe the weakest point of a data analysis agent: NL2SQL accuracy. The answer highlights the four improvement methods above, emphasizing schema clarity, few‑shot prompting, error‑feedback loops, and human verification for risky queries.

Key takeaway: a data analysis agent consists of intent recognition, NL2SQL‑based data retrieval, statistical/visual analysis, and natural‑language result presentation, with NL2SQL being the primary bottleneck.