Skip to content
← All workKirti KumarNext →

Product Design · take-home exercise · 7-day sprint

An AI agent that fixes your storefront while you watch

End-to-end design for an AI agent in the Shopify ecosystem — scan, diagnose and autonomously repair storefront issues, with configurable trust modes and a full audit trail.

My role
Sole designer
Context
Shopify ecosystem · take-home
Method
7-day end-to-end sprint
Output
Full product design + system

[ 01 · The brief ]

Autonomy is easy to build and hard to trust.

The brief was an agent that could find and fix storefront problems by itself. The design problem was not detection — it was consent. A merchant will not hand write-access to their live store to something they cannot supervise.

So the product had to make autonomy adjustable, and every action reversible and inspectable.

Landing and positioning
Diagnostic surface
Agent trust modes

[ 02 · Process ]

What I did

  • Mapped the merchant mental model around risk, revenue and reversibility
  • Designed three trust modes: suggest only, approve each fix, and autonomous within limits
  • Built a diagnostic surface that ranks issues by revenue impact, not severity score
  • Designed a full audit log so every autonomous action is attributable and undoable

“The unit of trust is not accuracy. It is reversibility.”

Design principle, ShopOS

[ 03 · Findings ]

The design position

Rather than asking merchants to trust the agent globally, the product lets them grant autonomy per category and per risk level, then widen it as the track record builds.

Every fix ships with a before and after, a plain-language reason, and a one-click revert. Trust is earned incrementally and visibly.

Before and after a fix
Audit log
7
Days from brief to full design
3
Configurable autonomy modes
100%
Actions logged and reversible

[ 04 · Outcome ]

What changed

The exercise produced a complete product design: positioning, diagnostic flow, agent configuration, audit trail and a component system. The core move was reframing an accuracy problem as a permissions problem.

Agent products fail on governance long before they fail on capability. Designing the leash is the design work.

Reflection
[ Next case ]

Time tracking people actually complete on time

Kirti Kumar © 2026Start a project →