Case Study

Enterprise HR Agentic Chatbot

Client: HTI Group Period: Apr 2025 – Dec 2025 Role: AI Engineer & LLMOps

TL;DR

Multi-agent HR chatbot built on LangGraph + LlamaIndex with the ReAct pattern, hierarchical semantic chunking, and hybrid retrieval (Vector + Full-text + Graph + Pandas). Migrated from local POC to production AWS infrastructure with Lambda, ECS Fargate, and distributed Langfuse observability. CI/CD via AWS CDK in TypeScript.

LangGraph LlamaIndex ReAct Pattern AWS Lambda ECS Fargate AWS CDK TypeScript Langfuse Qdrant Hierarchical Chunking

Problem

HTI Group needed a single chatbot that could carry the weight of fragmented HR institutional knowledge — and then survive the jump from quick demo to enterprise production.

  • Fragmented knowledge spread across Excel, Docs, Markdown, PowerPoint, PNG, and CSV sources.
  • Off-the-shelf RAG broke down on complex policy queries that crossed multiple sub-domains: leave, payroll, benefits, tooling.
  • POC-to-production was non-negotiable: the agent layer had to evolve to AWS without a rewrite.
  • Observability gap: multi-agent systems fail in subtle ways that unit tests miss.

Phase 1 — Proof of Concept

High-Level System Architecture

The foundation of the HR Chatbot. This diagram details the core interaction flow between the front-end user interfaces, the centralized API gateways, and the underlying AI knowledge engines before scaling up.

High-Level System Architecture

Multi-Agent Router (ReAct)

The brain of the chatbot. Using the ReAct methodology, queries are intercepted, understood, and dynamically routed to specialized sub-agents. This lets the bot handle complex HR policies, leave requests, and payroll inquiries with high accuracy — and makes each sub-agent independently testable and observable.

Multi-Agent Router

Data Ingestion Pipeline

A robust offline pipeline that automatically ingests, chunks, and maps unstructured HR documents into a high-dimensional vector space for semantic retrieval. Uses Hierarchical Semantic Chunking with Auto Merge, plus LLM Contextualization for child chunks so each retrieval unit carries its surrounding context.

Data Ingestion Pipeline

Phase 2 — Enterprise Production

AWS Cloud Migration

Challenge The POC ran on a single box. Enterprise needed 99.9% uptime, auto-scaling, VPC isolation, and a clean split between short-lived tool calls and long-running agent workers — without rewriting the agent layer.

Transitioning from local/on-prem POC to AWS. Leveraging fully managed services and container orchestration to reach enterprise targets: 99.9% uptime, auto-scaling, and security inside a VPC. Compute split between AWS Lambda (event handlers, short-lived tools) and ECS Fargate (long-running agent workers).

AWS Cloud Migration Architecture

Serverless Ingestion Pipeline

An alternative, highly optimized ingestion architecture using serverless cloud functions. Designed for event-driven document processing — minimizes idle compute cost and maximizes real-time scalability during data upload spikes. Useful when ingestion is bursty and unpredictable rather than steady-state.

Serverless Ingestion Pipeline

Outcome

  • Unified institutional knowledge across 6+ document formats behind one chatbot interface.
  • Production-ready on AWS with two deployment topologies (Lambda + ECS Fargate, plus a serverless ingestion alternative).
  • Hallucination-free outputs via Langfuse-traced agent trajectories and per-step evaluation.
  • Reusable agent and infra patterns that informed downstream projects at FPT Software.
  • Tool calling to external services (Discord, Google Calendar) integrated cleanly through the agent framework.