2026AI triage

PulseDesk

A support desk that reads its own inbox: comments go to a language model, and the ones that are real problems come back as tickets with a category, a priority and a summary.

Role: Backend and the model contract
Year: 2026
Built with: Java 21
Spring Boot
Spring Data JPA
Llama 3
H2

A Spring Boot service in Java 21 with one interesting seam in it: HuggingFaceService, which asks Llama-3-8B-Instruct four questions about a comment at once — is this actually a problem, what kind, how urgent, and what is it in one line — and hands back something the rest of the application can treat as data.

The whole design is that seam. A model will happily answer in prose, and prose is not a column, so the prompt asks for a fixed shape and the parser refuses anything else: BUG, FEATURE, BILLING, ACCOUNT or OTHER, and LOW, MEDIUM or HIGH. Nothing downstream knows an LLM was involved. CommentService sees a decision, TicketRepository sees a row.

A model that is allowed to be creative about its output format is a model you will be parsing forever.

The layers are ordinary on purpose — controllers, services, repositories, JPA entities — because the unfamiliar part is the model call, and the way to keep an unfamiliar part reviewable is to give it a boring frame. Storage is H2 in memory: this is a demonstrator, and a demonstrator that needs a database installed before it will start is a demonstrator nobody runs.

There is a small web UI over it — tickets on the left, incoming comments on the right — so the flow can be watched rather than described. Written against an IBM internship exercise brief.

Source