Agentic AI & LangGraph🏆 1st Place Winner // Future Talent & Innovation Team Award1st Place Winner (20+ Teams)

LangGraph Multi-Agent Enterprise Insights Platform

// Autonomous multi-agent orchestration for enterprise dataset profiling, SQL generation, visual charts, and executive synthesis.

Organization / ContextInfosys BPM — AI Centre of Excellence
My RoleLead AI Engineer (Led team of 3 engineers)
TimelineDec 2024 — Present
LangGraphPythonAzure AI ServicesOpenAI APIFastAPIPandasMatplotlib

1. Executive Overview

Enterprise business leaders and operations managers frequently need rapid answers from complex, multi-table transactional databases. Traditional BI workflows require dedicated data analysts to write SQL queries, extract CSVs, clean anomalies, compute statistical distributions in Python, and manually build presentation decks.

At the Infosys BPM Internal AI Hackathon, I led a high-performing team of 3 engineers to architect and build the Multi-Agent Enterprise Insights Platform. Powered by LangGraph, our system eliminates analytical bottlenecks by orchestrating specialized AI subagents that collaboratively turn natural language business inquiries into verified data transformations, publication-ready charts, and executive decision memos in seconds.

System Topology & LangGraph State Machine

StateGraph
STAGE 01: SUPERVISOR AGENT (Intent & Task Routing)

Analyzes natural language query, validates schema metadata, identifies required tables, and constructs execution sub-goals.

STAGE 02: DATA PROFILER & TRANSFORMER (SQL & Pandas Engine)

Executes verified read-only SQL queries, performs dynamic aggregations, filters statistical anomalies, and outputs clean tabular data frames.

STAGE 03: VISUAL CHART GENERATOR (Matplotlib / Seaborn)

Dynamically writes and validates Python plotting code in an isolated execution sandbox, generating publication-ready visual distributions.

STAGE 04: EXECUTIVE SYNTHESIS & CHATBOT INTERACTION

Synthesizes numeric trends into concise executive bullet points with verifiable source citations and contextual conversational follow-up capabilities.

2. Technical Challenges & Architectural Innovations

Challenge: Hallucinated SQL Queries & Schema Mismatches

Solution: We implemented a two-tier schema validation loop in LangGraph. Before query dispatch, a schema verification tool checks generated table aliases against live database metadata, automatically returning syntax errors to the SQL agent for self-correction.

Challenge: Stateful Multi-Turn Conversational Memory

Solution: Utilizing LangGraph's persistent state checkpointers (MemorySaver / PostgresSaver), the system retains dataset contexts across conversational turns, allowing users to drill down into specific chart cohorts without re-executing heavy ETL pipelines.

🏆

Hackathon Victory & Business Impact

Our team presented this working system to a panel of principal architects and business leaders at Infosys BPM. Competing against 20+ engineering teams, our platform secured 1st Place and was honored with the Future Talent & Innovation Team Award.

  • 90% Reduction in time-to-insight for operational queries.
  • Zero Hallucination on structured SQL metrics via schema sandboxing.
  • Transitioned from hackathon prototype to production incubation in the AI CoE.
Copied to clipboard!