AI Chat for Analytics

The Problem

The client manages a global portfolio of commercial real estate assets worth billions. Their business users (asset managers, analysts, investor relations) constantly needed data insights:

• "What's the average occupancy across our NYC office properties?"
• "Show me properties with leases expiring in the next 18 months"
• "Compare NOI trends across our retail vs. industrial portfolio"

Every question required a ticket to the data engineering team. Turnaround was days, sometimes weeks. By the time analysts got answers, they had new questions. The bottleneck was killing productivity.

The Solution

I built a conversational AI interface that lets business users query their property data in natural language. No SQL required, no waiting for data engineering.

How It Works

Natural Language to Analysis

Users type questions in plain English. The system translates their intent into Python/SQL queries, executes against the property database, and returns visualizations and insights.

Self-Correcting Agent

Here's where it gets interesting. LLMs make mistakes. They generate queries with syntax errors, reference wrong columns, or misunderstand the schema. I engineered a self-correcting agent that:

1. Generates initial query based on user question 2. Executes and catches any errors 3. Analyzes the error message and schema context 4. Regenerates a corrected query 5. Repeats until success (typically 1-3 iterations)

Few-Shot Learning

When a query succeeds, the system captures the question-to-query mapping as a few-shot example. Over time, the system builds a library of successful patterns specific to this client's data, improving accuracy and reducing iterations.

Technical Architecture

Stack:

• React frontend with conversational UI
• Flask backend orchestrating the AI agent
• Azure SQL Database holding property data (4,000+ rows across multiple tables)
• Azure App Service for deployment
• Okta SSO for enterprise authentication

Key Design Decisions:

• Agentic Architecture: The LLM doesn't just generate SQL. It reasons about the data model, handles errors, and iterates toward correct answers
• Schema Context Injection: Full database schema and sample data injected into prompts for grounding
• Guardrails: Read-only access, query timeouts, and result size limits prevent runaway queries
• Explanation Layer: System explains its reasoning, so users understand (and can correct) the analysis

The 6-Week Sprint

This project had aggressive timelines. The client needed a working demo for an upcoming investor meeting. We delivered:

Week 1-2: Schema analysis, data profiling, initial agent architecture

Week 3-4: Core conversational interface, self-correction logic, error handling

Week 5: SSO integration, security review, performance optimization

Week 6: User testing, refinements, production deployment

Results

The platform eliminated the data engineering bottleneck entirely. Business users now get answers in seconds instead of days. The client was so impressed they integrated the technology into their own B2B offering for their investors.

Key wins:

• Hours to seconds: Analysts get insights immediately
• Self-service analytics: No more tickets to data engineering
• B2B integration: Client monetizing the capability with their own customers
• 6-week delivery: From kickoff to production in aggressive timeline