Skip to content

AI Agents Platform

Web appShipped work

An internal platform that brings every business unit's AI agents into one place and puts them behind a single conversational chat.

Stack

  • Next.js
  • React
  • React Query
  • shadcn/ui
  • Supabase
  • n8n
The platform's workspace directory, one workspace per business unit

Snapshot

  • Role — Frontend Engineer: the whole interface, the frontend architecture, the n8n integration, deployment, and proposing what to build next.
  • Scope — Internal tool in production: 50+ business units, 80+ agents, 100+ users.
  • Outcome — The 80+ agents that used to live scattered across n8n are now inventoried in one place and used through a single chat.

The problem

Every business unit built its own agents in n8n, its own way. Nobody had the full list: not which agents already existed, not which ones each unit was using, not which ones were missing. The cost wasn't just untidiness — a unit had no way of knowing whether what it needed had already been built somewhere else in the company.

Key decisions

  • shadcn/ui instead of building the component system — the idea had to be seen working before it could be argued about, and with the components already solved the interface moved at the speed of the product.
  • Supabase because the company already paid for it — availability, not a technical fit: using what was already there skipped an infrastructure debate that would have delayed the MVP.
  • n8n's REST API instead of reimplementing agent execution — the team already lived in n8n and knew how to build there; reimplementing that layer would have meant teaching them another tool on top of maintaining it. The tradeoff: the platform is tied to n8n and to how each unit wired its own workflow.
  • No mobile in the MVP — validating the idea with real users came first. The tradeoff: the platform was born desktop-only.

Outcome

  • 50+ business units with their agents centralized, 80+ agents inventoried, and 100+ people using them through the platform.
  • Still under active development: the MVP didn't close the project, it opened it.
Dashboard ranking agents by number of runs, next to a list of users with no recent activity
The dashboard answers what nobody owned before: which agents actually get used, by how many people, and who stopped showing up.
Chat with a creative agent returning an edited image and a link to the folder holding the result
The unit asks for the change in plain language and gets the deliverable back; underneath, it runs the same n8n workflow as always.