Orion
2025
Orion: AI agent operations console
An operations console for teams running AI agents in production. Monitor every agent, replay any run step by step, approve risky actions before they execute, and cap spend, so autonomy stays supervised without slowing anyone down.
Overview
Orion Overview
Orion is a control panel for companies that use AI agents to do work, like answering support tickets, issuing refunds or reviewing contracts. It shows how all the agents are doing, flags what's going wrong, and lets a person step in when an agent needs approval.This project is protected under a non-disclosure agreement and cannot be shared publicly.
The goal was to create a product that helps users discover new artists, explore collections, and enjoy a more engaging listening journey without overwhelming navigation or complexity.
If you’d like to learn more about the project, process, or my role, feel free to get in touch.
Define
What problem does it solve?
When a company runs many AI agents, it's hard to know what they're doing. Problems like an agent failing, getting worse over time or spending too much money are noticed late, and the details are scattered across logs. Risky actions, like a large refund, have no clear point where a person checks them first. It's also hard to see which agent is costing how much. As a result, teams spend their time reacting to problems instead of preventing them.
Ideate
Information Architecture
Exploring interface directions
Designs
Dashboard experience, Agents, Runs, Approvals
Control Room (dashboard): Shows fleet-wide health at a glance. It has KPIs (active agents, task success, escalations, spend), run volume by outcome, fleet health, alert cards for active issues, and a live run stream. It answers "is everything okay right now, and what needs me first?"
A filterable list of every agent with status, 24h volume, success rate, latency and version. A side panel for the selected agent shows its incident, tool reliability and guardrails. It lets you compare agents and spot the weak one without opening each.
A searchable table of all runs with filters (status, agent, failing tool, cost) and bulk actions: retry, group as incident, assign. It turns hundreds of failures into one actionable set, such as every stripe.refund timeout.
Cost & Usage: Spend trend, spend split by model, budget tracking with a projected cap date, and cost by agent. It shows who is spending what, so budget overruns are caught before they happen.
Design System
Design System
