Nomin GPT, an internal AI assistant for ~6,000 employees
- ~6,000
- employees with access in Teams
- 755
- regulations searchable (11,234 pages)
- 3
- agents (Triage, HR, IT)
TL;DR
- I proposed an internal AI assistant, built the first version on Copilot Studio, and project-managed it to a company-wide rollout in Teams (~6,000 employees).
- When 755 regulations (11,234 pages) outgrew Copilot Studio, I led the move to a custom 3-agent system with the E-Commerce engineering team.
- The custom system answers HR and IT questions from the regulations, looks up payroll, attendance and leave in Odoo ERP, and opens IT tickets on its own.
- My role
- Initiator and project manager; built v1 myself, led the custom version with the E-Commerce engineering team
- Year
- 2025–2026
- Stack
- Copilot Studio · Microsoft Teams · FastAPI · OpenAI Agents SDK · Gemini · Typesense · Odoo ERP
Context
Nomin Holding is one of Mongolia’s largest retail and trade conglomerates, with more than 6,000 employees (company figure, 2025). I joined its Digital Transformation Office in October 2025 as a Business Process Automation Engineer and Project Manager.
Two kinds of employee questions were the starting point: IT support, and HR and labour rules. The answers to both already existed, in the company’s regulations and internal systems.
Problem and scope
The assistant was my proposal. I suggested the idea and scoped a first version that could prove value quickly inside the tool everyone already used, Microsoft Teams:
- an IT-support helpdesk bot, and
- an HR and labour-rules bot,
both built on Copilot Studio, so the company could test the idea without new infrastructure.
Approach and key decisions
Start on the platform the company already had. Version 1 ran on Copilot Studio inside Teams. That kept the first release inside the existing Microsoft 365 setup and let employees try it where they already work.
Move to a custom system only when the platform stopped fitting. Once the regulation corpus reached 755 regulations and 11,234 pages, it outgrew what Copilot Studio could handle well. I led the move to a custom multi-agent system, built together with the E-Commerce engineering team:
- 3 agents, Triage, HR and IT, on FastAPI with the OpenAI Agents SDK and Gemini. Triage routes each question to the right specialist.
- Typesense retrieval (RAG) over the full set of regulations, so HR answers come from the company’s own documents.
- Odoo ERP integration for payroll, attendance and leave, so an employee can ask about their own records, not only about the rules.
- Automatic IT tickets, so an IT conversation can end in a ticket without a separate form.
What broke
The first version hit a ceiling. Copilot Studio was a fast way to prove the idea, but the regulation corpus (755 regulations, 11,234 pages) outgrew it. I took that limit as the signal to change architecture, and led the move to a custom system with the E-Commerce engineering team.
Result
- Nomin GPT is rolled out in Microsoft Teams to about 6,000 employees.
- HR questions are answered from 755 regulations (11,234 pages) through retrieval (RAG).
- Payroll, attendance and leave lookups come straight from Odoo ERP, and IT conversations can create tickets automatically.
My role compared with the team’s
- Mine: I proposed the idea, built v1 on Copilot Studio (both bots), project-managed the whole effort, and led the move to the custom system.
- The E-Commerce engineering team’s: building the custom multi-agent system with me (FastAPI, Agents SDK, RAG, ERP integration).
Stack
Copilot Studio · Microsoft Teams · FastAPI · OpenAI Agents SDK · Gemini · Typesense (RAG) · Odoo ERP
Proof
Internal system, so there is no public demo. I am happy to walk through the architecture and the rollout in an interview, and a reference is available on request.