← Back to Projects

EnterpriseSG · Aug 2022 – Feb 2025

Genie + Oracle

800+ non-technical users querying data via natural language. Agency's first LLM applications

Spot me teaching SQL! (LinkedIn) →

800+

Active users

1,060

Prompts processed

36%

SQL query increase

4/5

User satisfaction

The Problem

EnterpriseSG is a Singapore government agency with 800+ employees. Institutional knowledge was locked in long reports, databases, and meeting minutes. Officers spent days searching for insights and there were no AI tools in the organization.

What I Built

Two complementary LLM applications:

Genie

Text-to-SQL tool enabling natural language database queries. Reduced SQL writing time from 15 minutes to 30 seconds.

Oracle

LLM chatbot for searching and summarizing meeting insights. Reduced search time from 3 days to 4 hours.

I also built an automated evaluation pipeline using LangChain to assess response quality, designed template/default prompt features (accounting for ~15% of all queries), and used KMeans clustering to analyze usage patterns and improve the product.

Shipping AI in Government

I created EnterpriseSG's first internal LLM policy, conducted a comprehensive risk assessment, and secured buy-in across 8 internal and external stakeholder platforms. I ran 10+ user testing sessions and sharing sessions reaching 300+ users, and gathered data from a 100+ participant survey to shape the roadmap.

Data-Driven Iteration

To diagnose declining usage, I used KMeans clustering to segment query types, identified that users were trying to ask quantitative questions Oracle wasn't designed for, and built a sentiment extraction feature that processed meeting minutes at scale to handle that need.

Tech Stack

PythonLangChainOpenAIReactJSFlaskSQLStreamlitKMeans