Oversee

Workshop management system where organizations track vehicle repairs, assign technicians, and monitor performance. Users create reports with registration numbers, problem descriptions, categories, and time tracking. An org-scoped AI assistant powered by Z.AI GLM-4.7 handles multi-step workflows — parsing Excel uploads into reports, querying the database, generating charts, and bulk-updating records — with human-in-the-loop approval before any destructive action. The whole interface is Swedish-first, built for field teams who may not be familiar with AI tools.

• –Role-based multi-tenancy: owners, admins, members, and technicians each see different data and controls, enforced by Supabase RLS
• –AI assistant with 10+ tools for database queries, report CRUD, Excel parsing, chart generation, and web search — all scoped to the active organization
• –Human-in-the-loop approval dialogs before the AI creates, updates, or deletes reports
• –Report management with registration numbers, categories, technician assignment, time tracking, and bulk operations via drag-reorderable data tables
• –Dashboard analytics: KPI cards, daily/weekly/monthly trend charts, category distribution, technician performance bars, and top reported issues
• –Vehicle inspection tracking (Bilkollen) for recurring check-ups alongside ad-hoc repair reports
• –Invitation system with join requests, org creation flow, and onboarding for new users

• –Z.AI's GLM-4.7 only supports Chat API, rejects system role messages, and doesn't handle images — so the integration needed workarounds for message formatting and graceful degradation when users attach files
• –Multi-step AI workflows like Excel import (fetch technician names → category names → parse → create reports) had to preserve context across 5+ tool calls without blowing up the token budget
• –Technicians should only see their own reports while admins see everything — getting RLS policies right for that asymmetry across all query paths took careful testing
• –Base64 images from screenshots can exceed Vercel's 8MB request limit, so the client compresses them iteratively before sending

System prompt gets prepended to the first user message to work around GLM's missing system role. AI tools use names_only flags and pagination to keep token usage low during multi-step operations. RLS policies enforce org + role scoping at the database level, and server actions validate membership before any mutation. Client-side Canvas API compression reduces image payloads by 60-80% with iterative quality reduction. Rate limiting (30 req/hour per user per org) prevents abuse without blocking normal usage.

Used in production by a workshop organization to track repairs and technician workload. The AI assistant handles the bulk of data entry — field technicians snap a photo or paste an Excel export and the assistant turns it into structured reports after confirmation. What started as a KPI dashboard grew into the primary tool for day-to-day operations.