HR & BPO · Philippines, USA

AI assistants for an HR and BPO platform

LLM features added to a working enterprise platform: chat, ticket triage, insights and dashboards

Industry: HR & BPO
Geography: Philippines, USA
Type: Web, AI
AI in production: Since July 2024
AI features: 6
Technologies: Python
OpenAI API
LLM
RAG
Vector search
Redis
AWS

AI assistants for an HR and BPO platform

Project overview

The client is an outsourcing company with its own platform for HR, operations and client projects. Employees and client managers use it every day to find documents, raise tickets and track work.

Since 2023 the company has been adding AI features to the platform so that people spend less time searching for answers and routing requests.

What we did

Our team works on the platform alongside the client's own engineers. On the AI side we built and ran the backend services and the infrastructure behind them.

2023: proof of concept of a chat assistant that answers questions from the company's own documents
July 2024: the assistant in production, with retrieval from a vector index and a cache for repeated questions
An operations chat and AI insights that summarize activity for managers
A dashboard generator that builds a chart from a plain-language request
Ticket triage on a self-hosted open-source LLM, so ticket text stays inside the company's cloud
A document review assistant for one of the client's customers

Key results

AI chat assistant in production since July 2024

Six AI features running on the same platform

Ticket triage on a self-hosted model, with no ticket data sent to an outside API

The chat service runs in its own scaling group, so a spike in AI requests does not slow down the rest of the platform

Load was the hardest part. Each chat request used a lot of memory, and before an important client demo the service became unstable. We moved it to a separate scaling group and added caching.

Triage suggests a category and owner for a new ticket as soon as it arrives, so routing that used to wait in a queue for hours now takes minutes.

Frequently asked questions

Why run ticket triage on a self-hosted model?

Tickets often contain personal and client data. A self-hosted open-source model keeps that text inside the company's own cloud account, and for a narrow task like choosing a category it is accurate enough and cheaper per request.

How does the chat assistant find answers?

It searches a vector index built from the company's own documents and passes the relevant passages to the LLM with the question. Answers to repeated questions come from a cache, which keeps response times and model costs down.