AI Workflow Dashboard.

Overview

An internal-facing dashboard for operators running AI-driven workflows. Hundreds of in-flight tasks, tool calls, partial states, and human-in-the-loop checkpoints — all needing to be legible at a glance.

Problem

The previous interface buried state. Operators were missing failures, retrying things that were already running, and losing trust in the system. The dashboard was busy without being useful.

Constraints

• Real-time updates over many concurrent jobs without melting the browser
• Operators needed to act quickly — UX had to bias toward decisions, not exploration
• No room for a redesign sprint — work had to ship in slices behind a flag

Approach

Rebuild around “state at a glance.” Define every job state explicitly. Make the most operator-relevant signal the loudest. Push detail off the main view into a focused side panel. Then layer streaming updates so the screen reflects truth, not a stale snapshot.

Implementation

Next.js with server components for the heavy lists, client components for the live state. Tanstack Query plus Supabase realtime for the streaming layer. A small but strict design system for status pills, timing, and severity so nothing was ambiguous.

Result

Operator action time dropped meaningfully. Repeated retries fell sharply. The team started trusting the dashboard as the source of truth instead of cross-checking logs. The interface stopped being an obstacle.

Reflection

The clearest dashboards aren’t the prettiest — they’re the ones where the most-needed signal is impossible to miss. Most of this work was deletion: cutting columns, charts, and labels until what was left earned its space.