MDA AI Helpdesk Assistant
A retrieval-augmented IT helpdesk prototype for the Mississippi Development Authority, built with the Mississippi AI Innovation Hub and selected at the statewide AI showcase. It was a prototype and was not deployed.
A retrieval-augmented IT helpdesk assistant for a Mississippi state agency of about 250 employees, built by a team of three through the Mississippi AI Innovation Hub. It was chosen at the statewide AI showcase in Jackson. It remains a prototype and was not deployed to production.
The problem
Helpdesk answers live in unstructured documents, and a wrong answer in a public-sector setting has real consequences. The assistant needed to answer accurately when it could, and hand off to a person when it could not.
How it works
Documents are stored in S3 and embedded with Amazon Titan. Retrieval runs as k-NN search on Amazon OpenSearch Serverless, and answers are generated with Amazon Nova Lite behind a FastAPI service running in Docker.
My part was the retrieval design, the prompt engineering, and the confidence scoring. When confidence is low, the assistant escalates to a human instead of guessing, and it suppresses duplicate tickets.