This thesis describes the design and evaluation of an agentic system for managing
applications and reporting in a small non-governmental organization, the Global Shapers
Institute, which works in youth empowerment. With a limited budget, 200 applications had to
be processed and biweekly reports for a grant compiled.
The work the team had previously done by hand was divided into two layers. Steps that are
always the same run as automations: application confirmation, notice of a submitted form and
reminders before events. Steps that require judgement are carried out by two agents that
reason with the help of large language models. The first compiles the reports for the grant, the
second handles communication with the candidates.
The solutions that lasted are small, save the team manual work and grew step by step out of
tasks that were previously done by hand. The thesis also describes the attempts that failed and
draws from them when an agent is worth building at all and how to build it so that it stays
understandable and maintainable. It shows how presenting an agent with a diagram and a
legend, together with a precise description of its task, makes it possible to review every agent
in the same way.
|
