AI Product & Behavior Designer

The space where product intent becomes model behavior.

I design AI agents that know when to act, ask, stop, and hand off. Then I build the prototype, write the evals, and turn product decisions into model behavior teams can test.

I started in language-model training and NLU, moved through production voice and chat assistants, then into LLM and agentic systems. My edge is translating messy customer problems into behaviors an AI system can execute, evaluate, and safely hand off.

Where: Intuit · Credit Karma · Sam’s Club · Apple Siri · Meta

View selected work

Selected work

Six cases, from LLM classifiers to agent skills.

Every cover is a designed receipt. The real artifact lives inside each case.

Intuit Intelligence · Billing skill

Every charge,
explained.

grounded in the customer’s own data

personalized responses explainability first

Featured · Agentic era

“What is this charge?” is the most common billing question, and the hardest to answer well.

A skill that grounds every answer in the customer’s own account: personalized, explainable, and honest enough to intercept a cancel before it happens.

Intuit Intelligence · Quality program

3×

faster to golden guidance

Quality program · Agentic era

How do you grade an agent when there’s no answer key?

You write one, at scale: generator prompts draft, twelve QA checks screen every batch, and experts keep the final say.

Intuit Intelligence · Payroll tax skill

what good looks like

personalized, jurisdiction by jurisdiction

Payroll tax skill · Agentic era

Payroll tax guidance lived in a few experts’ heads. That doesn’t scale.

So define what a good diagnosis looks like, personalize it jurisdiction by jurisdiction, and grade the live skill against it.

QuickBooks Voice Assistant

+3.9pt

containment lift, shipped

Voice · LLM era

On a phone line, there’s no screen to fall back on.

Every word is the interface: what the model can say in seconds, when to clarify, when to hand off. Shipped a +3.9pt containment lift.

Vapi prototype · Voice troubleshooting

fixed on the call

a live walkthrough, end to end

Voice troubleshooting POC · LLM era, 2024

Self-help sends the answer. This prototype stayed until it was fixed.

System instructions, grounded product content, and a live voice loop. Hear the model clarify, pace itself, admit a limit, and recover.

▶︎ Hear the session

Intuit · Cancellation saves

96.6%

routing precision, in evaluation

Cancellation saves · LLM era

Not everyone who clicks “cancel” actually wants to leave.

A classifier accurate enough to route the truly savable to a specialist, and give everyone else a clean goodbye.

Handoff

If your team is deciding how an agent should behave, I can help.

simonermuro@gmail.com LinkedIn Request resume

About the artifacts and how this site was made: colophon.