Building Conversational AI & NL2SQL Engine from 0-1
Full-stack analytics application modernizing conversational AI pipelines and automated NL2SQL query synthesis for enterprise data lakes.
Lead Engineer
Core Builder
2024 — 2026
Python FastAPI, Azure AI Search, PostgreSQL
The Friction Point & Opportunity
Enterprise stakeholders struggle with manual SQL query formulation against massive data lakes, leading to slow reporting cycles and high operational bottlenecks across data teams.
System Design & Implementation
Built a production RAG pipeline leveraging Azure AI Search, FastAPI async orchestration, LangChain query decomposition, and vector caching.
SYSTEM TOPOLOGY & DATA PIPELINE
### Pipeline Architecture
```mermaid
flowchart LR
User[Enterprise User Query] --> API[FastAPI Gateway]
API --> Intent[Intent & Schema Parser]
Intent --> Vector[(Azure AI Vector Search)]
Intent --> SQLGen[NL2SQL LLM Engine]
SQLGen --> Validator[SQL Guardrail & AST Validator]
Validator --> DB[(Enterprise Data Warehouse)]
DB --> Aggregator[Response Synthesizer]
Aggregator --> User
```Key Capabilities & Engineering Depth
Sub-3s NL2SQL Generation
Translates natural language questions directly into performant SQL queries with automated index awareness.
Hybrid Search & Semantic RAG
Combines sparse BM25 and dense vector retrieval over 2,000+ internal policy documents.
Tactile Micro-interactions & Usability
Interfaces should feel alive, responsive, and predictable. Every touchpoint was tested against real user mental models to ensure zero confusion during multi-step tasks.
Measurable Outcomes & Production Resilience
Cut report generation time from 10s to <3s for 2,000+ active enterprise users, reducing support escalations by 40%.
Key Learnings & Future Roadmap
Building software from 0 to 1 reinforced the principle that great engineering is about discipline, simplicity, and continuous feedback loops with real users.