AI
CompletedInternal AI Assistant
An assistant that helps colleagues understand an internal application and turn ideas into clear change requests, with people approving every request before it reaches IT.
Cannot change systems6/6
Any real change stays with people.
The problem
Colleagues often have good ideas for improving internal software, but turning them into a clear change request takes time.
An AI tool connected to a business system is only acceptable if it cannot damage that system.
My contribution
- Designed the operating model: read, analyze, propose, prototype and document, with no path to write, deploy, migrate or delete.
- Built search over documentation, so that answers come with evidence.
- Built an approval workflow, where people decide what reaches IT.
- Made model providers configurable, including an on-premises option.
Technical architecture
- Interface
- Chat, prototypes and approval workflow
- Application
- API service
- Knowledge
- Searchable index, kept apart from business data
- Models
- Configurable hosted or on-premises models
Technologies
- Python
- FastAPI
- React
- Claude API
- OpenAI API
- MCP
Highlights
- Reads and proposes, and does not change systems
- Approval by people, not by the model
- Answers come with evidence
Details are generalized to protect confidential information. Company systems, customers, figures, source code and screens are not shown.