AI Helpdesk Agentic Flow
Designed and built end to end by one engineer. Architecture, agent, backend, web app, tests, CI and infra.
An IT helpdesk where employees ask for system access in plain words. An AI agent works out what they need, a fixed policy engine decides, and the system grants it for a set time, sends it to human approvers, or blocks it. The model never changes anything itself, and every decision names its rule in an audit log.
- 42conversation evals, prompt injection attacks included
- 38MCP tools, each behind a role check
- 100%test coverage on the policy engine
What you'll see in the demo
- Press Sign in to the demo and pick Employee. You get your own made-up company, deleted after a day.
- Ask "Give me Internal Wiki viewer access". It is low risk, so the agent grants it at once.
- Ask "Give me Internal Wiki editor access". This one needs approval, so it goes to your manager.
- Sign in again as Approver. You are that manager, and you can approve or deny the request.
- Sign in as Approver and ask the chat to remove the Wiki editor access you just approved. A manager can remove access in their own team.
- Sign in as Employee and ask it to remove a colleague's access. It refuses, because an employee can only remove their own.
- Sign in as Admin to read the audit log of every decision.
After you try it, open the LLMOps Console demo to see the model calls it just made.
The code, the design and how to run it locally are in the helpdesk-agentic-flow repo on GitHub.
github.com/masudur-rahman-musa/helpdesk-agentic-flow ↗Skills used
Applied AI
- Agentic AI with LangGraph
- MCP tool server
- RAG pipeline
- Qdrant vector search
- Versioned prompts
- LLM guardrails
- Conversation evals
Backend
- Python (async)
- FastAPI
- SQLAlchemy
- Alembic
- PostgreSQL
- Redis
Security
- RBAC
- JWT auth
- Multi-tenant isolation
- Audit log
Frontend
- Next.js
- React
- TypeScript
- Tailwind
DevOps
- Docker Compose
- GitHub Actions
- Kubernetes (Kustomize)
Testing
- pytest
- Playwright