Runs each investigation as a LangGraph pipeline: request guard, permitted retrieval, operations data, decision signals, model routing, brief and work-order proposal.
Full stack · Guarded AI workflow
Service Desk
An AI-assisted tool for field service teams. It gathers the manuals, bulletins, service history and parts a case needs, drafts a sourced escalation brief, and never writes a work order without a person's approval.
- Role
- Designer and developer
- Organization
- Personal project
- Period
- 2026
- Code
- service-desk ↗
Launch film · 1:18. Every value on screen comes from a real run of the app.Open the film ↗
System architecture
Architecture diagram created for this projectView full image ↗
01 · Context & problem
Service Desk
When a packaging line stops for the third time in a month, the engineer needs the service manual, the firmware bulletin, the maintenance agreement, the service history and the parts stock. Service Desk turns one request into a reviewable escalation brief without letting a model decide permissions or act on its own.
02 · What I built
Screens requests with a local guard model (Laya) for jailbreaks, prompt injection, policy evasion and data exfiltration before anything is searched.
Retrieves only the documents the signed-in user may read, combining pgvector similarity with PostgreSQL full-text search through reciprocal rank fusion.
Checks every citation in the brief against the passages actually retrieved and removes anything unsupported.
Routes the report only to a model endpoint cleared for the most sensitive passage in the evidence.
Locks work orders behind approval: the API checks the case has not changed, writes the work order and records an audit event.
03 · Investigation flow
01Request guard
→02Permitted retrieval
→03Operations data
→04Brief + citations
→05Human approval
04 · Engineering decisions
Permissions stay in the API
Access is enforced on every list, read, search and approval; the model never decides who can see what.
People approve the writes
A model may propose a work order, but an approver executes it, so the system cannot act on its own.
The guard runs before generation
A request that fails the local check never reaches retrieval or the writing model.
Claims stay traceable
Each citation is checked against retrieved passages, and the report goes only to a model endpoint cleared for the evidence.
05 · Stack
Next.jsReactTypeScriptFastAPIPostgreSQLpgvectorLangGraphLayaDocker Compose
Project details are based on available project documentation and résumé information.
Contact
Building something that needs real engineering depth?
Open to full-stack, backend, platform, DevOps and AI-integration opportunities in Québec, Canada and remote teams.