Human-Centered AI · Agentic Systems · Enterprise Workflows
Your agent shouldn't get to make the decision it isn't allowed to get wrong.
I design the interface layer between autonomous AI systems and the people accountable for what they do — where the model proposes, a human decides, and the boundary between the two is never ambiguous.
The failure mode
The model can be genuinely good and the product can still be dangerous — because nobody designed where human judgment enters, or what happens when the agent is wrong. That's an interface problem, and it's the one enterprises are least prepared for.
What the system is good at
- Drafting and pattern-matching at a scale no team can match
- Handling the repetitive 80% of a workflow correctly and fast
- Surfacing anomalies a tired human would miss
What still needs a human
- Anything with legal, financial, or safety consequences it can't undo
- Judgment calls with no clean precedent in training data
- Accountability — someone has to be answerable
What I design
Three things every agentic product needs, and most ship without.
01 / CONTROL
Load-bearing override points
Approval queues and kill switches placed where a wrong call is expensive — stopping the agent is one click, not a ticket.
02 / TRUST
Interfaces that show their work
Confidence indicators and reasoning traces surfaced in the UI, so "trust this" vs "check this" takes seconds.
03 / CONTAINMENT
Privacy and blast-radius by design
Scoping what an agent can see and touch before it acts, with an audit trail that survives a compliance review.
Before you email
A few things that usually come up before someone books a review.