Case Study
HRMS Chatbot — HR Assistant for Enterprises
Conversational HR services inside Slack and Teams, wired to a live HRMS portal
Demo video
What it is
The HRMS Chatbot is an enterprise assistant that lets employees, managers and HR teams handle day-to-day HR tasks conversationally: applying for leave, checking attendance, and querying policies. It integrates with the Keka HR portal on the backend and meets users where they already work — Microsoft Teams and Slack.
The problem
HR self-service portals are where productivity goes to die: logins, menus, forms and PDF policies for questions that take ten seconds to ask. Employees avoided them, HR teams drowned in repetitive tickets, and managers couldn't get quick answers about leave balances mid-meeting. The goal was to collapse all of that into a chat message.
How it works
Bot Framework + Azure infrastructure
The bot is built on Microsoft Bot Builder and hosted as an Azure Bot Service, which gives first-class Microsoft Teams channel support — critical for an enterprise audience — with Slack as an additional channel.
Intent routing with LangChain
Free-text messages are routed through a LangChain pipeline that classifies intent (leave application, balance query, attendance, policy question), extracts the relevant parameters, and dispatches to the right handler with validation before anything touches the HRMS.
Keka portal integration
A FastAPI middleware layer talks to the Keka HR portal's APIs: fetching balances, submitting leave requests and reading attendance records. Mutations go through validation and confirmation steps in the chat, so a leave application is reviewed in-conversation before it is submitted.
Policy questions via Chroma DB
HR policy documents are embedded into Chroma DB, so questions like "what's the carry-forward policy?" get grounded, source-backed answers instead of the model improvising company policy.
Challenges & solutions
LLMs hallucinate HR answers — and a wrong leave balance or made-up policy is a genuinely bad failure mode in an enterprise.
Strict separation: anything backed by data (balances, attendance, applications) goes through deterministic API calls with the LLM only handling understanding and formatting; policy questions are grounded via retrieval over the actual documents with citations.
Conversational applications need multi-turn state — "apply for leave" then "make it from Monday" must merge into one request.
Slot-filling conversation state per user/session in the Bot Framework dialog stack, so parameters accumulate across messages and the bot asks only for what's missing.
What I learned
- In enterprise settings, the LLM is the interface layer, not the data layer — routing and validation belong in deterministic code.
- Meeting users inside Teams/Slack matters more than any feature: adoption follows zero new apps to open.
- Confirmation flows for write-actions are non-negotiable when the bot can change real HR data.