I take AI products from zero to launch at Google scale, then use evals and customer feedback to make them work.

I lead Ask Google Help, Google's customer support AI agent for Google Store, Fi, and Pixel, and define the agentic actions framework (autonomous, human-in-the-loop, and consent-based) for support across Google products.

Previously, I spearheaded Edit with AI on YouTube Shorts, a generative video creation feature built with Google DeepMind on Gemini and shipped to millions of creators. I owned it from strategy and definition through execution and launch. Before moving into product management, I was a technical program manager at Instagram, Lyft, and Amazon, and a software engineer at FedEx.

My focus is product strategy for LLM-powered products, evaluation of nondeterministic models, data-driven iteration, and cross-functional leadership across research, engineering, design, legal, and operations.

0-to-1 product strategy
Product definition
LLM evaluation
Agentic systems
Trust & safety
Cross-functional leadership

View projects ↓

Experience

strategy → definition → execution → launch

Google · Current

Ask Google Help

Live for Google Store, Fi, and Pixel customers

Strategy

Own the agentic actions strategy for support across Google products, setting when agents act autonomously, with human review, or with user consent.

Definition

Architected the consolidation of fragmented troubleshooting systems into one agent orchestration layer. Defined how sensitive legal, safety, and privacy issues are routed to people.

Execution

Defined the quality bar for every answer and optimized against it to resolve more issues without a human handoff, while keeping high-risk cases with people.

Launch

Scaled the agent to handle a large share of troubleshooting, delivering higher customer satisfaction and stronger retention of users switching to Pixel.

Read the story →

YouTube · with Google DeepMind

Edit with AI

Shipped to millions of YouTube creators

Strategy

Spearheaded generative creation for Shorts to solve creators' blank-slate problem. Led the decision to build with Google DeepMind on Gemini rather than partner with an outside tool.

Definition

Defined the PRD and UX flows from creator research, covering story assembly, music, voiceover, and transitions. Defined model quality requirements with research engineers.

Execution

Prioritized speed over peak model quality and architected an asynchronous flow so creators never waited. Optimized quality through side-by-side evaluations.

Launch

Launched to millions of creators, shipping AI voiceover in multiple languages and AI-generated effects and transitions, and accelerated Shorts creation.

Read the story →

Earlier

Instagram · Lyft · Amazon · FedEx

B.S. Computer Science, UT Dallas · MBA, UCLA Anderson

Led cross-functional programs as a technical program manager at Instagram, Lyft, and Amazon, including defining Instagram's combined affiliate and branded content monetization. Began my career as a software engineer at FedEx.

Independent projects

built and shipped end to end

Cycling · crash intelligence

PeloRisk

Rates every Tour de France stage for crash risk, combining terrain, weather, peloton dynamics, and historical crash data. In outreach to professional teams.

LiveRisk scoring

Running

Sole Ally

Scan a shoe sole and get a tread-wear score, with the signals behind it shown. Runs fully on-device, so photos never leave the phone.

On-device ML

Pet health

Pawctionary

Ask about your dog's health and get answers grounded in veterinary resources.

LiveRAG

Product principles

applied across my work

Calibrated confidence

Products should communicate uncertainty explicitly. A clear "not sure" builds more trust than a confident error.

Safety and privacy by design

Safety and privacy are requirements set at the start, from on-device inference to strict data retention.

Prototype-driven discovery

I build working prototypes to test assumptions before writing a spec, which shortens the path from idea to evidence.

← Back to experience

YouTube Shorts · with Google DeepMind

Edit with AI

Taking generative video creation from a creator problem to launch.

1 · Problem

Many creators, especially new ones, struggled to get started. Facing a blank editor, they often didn't create at all. I led the discussion on how to lower that barrier through inspiration, ease of use, or both.

2 · Decision

We chose to redesign the creation flow around generative AI. The key call was the model: an outside tool would have integrated faster, but brought legal and privacy complexity. I led the build-versus-partner analysis, and we chose to build on Gemini with Google DeepMind, accepting a slower start.

Research with creators set the scope: understand the footage, assemble it into a coherent story, and add the finishing touches creators spend the most time on, like music, voiceover, and transitions.

3 · Execution

I wrote the product requirements and built the user flows with design. With research engineers, I defined what good editing looks like across common video styles, which shaped how the model was tuned.

Latency over peak quality. We chose a faster model over a higher-quality one and designed an asynchronous flow so creators never waited on generation.
Safety built in from the start. We tuned safety checks to catch harmful content reliably without slowing uploads, and set strict limits on how long user data is kept.
Speed over flexibility for voiceover. We pre-generated a small set of fine-tuned voices instead of generating them live, to keep the experience fast.
Inspiration that fits each region. I led an operations workstream to surface example videos and music that were relevant, fresh, and remixable in each market.

4 · Evidence

Side-by-side evaluations compared the model's edits against creator-made ones on coherence and quality, and we iterated on the gaps before launch. Music synced to the beat also improved creation beyond the AI flow.

Launched to creators and increased Shorts creation.

5 · Learning

Expanding AI voiceover to new languages showed that quality doesn't transfer automatically. Each language needs its own evaluation, not a copy of the English one.

← Back to experience

Google · Store, Fi, and Pixel

Ask Google Help

Rebuilding device troubleshooting around one AI agent, and defining how agents act for users.

1 · Problem

Troubleshooting help was spread across several separate systems, each handling part of the journey. At the same time, agents were gaining the ability to take actions, with no shared rule for when an agent may act on a user's behalf.

2 · Decision

We consolidated everything into one orchestration layer that sends troubleshooting to Ask Google Help. I also defined the agentic actions framework, now applied across Google products for support cases: when an agent acts autonomously, when a human stays in the loop, and when it needs the user's consent.

3 · Execution

I defined how response quality is judged, so every answer is clear, accurate, and leads to a solution. We kept improving quality to resolve more issues without handing off to a human agent.

I also defined how sensitive issues, such as legal, safety, or privacy concerns, are identified and routed to people.

4 · Evidence

Higher customer satisfaction, more troubleshooting handled by the agent, and better retention among users switching to Pixel.

5 · Learning

Deciding when an agent should hand off is a trade-off between protecting support teams from false alarms and making sure sensitive issues are never missed. We started conservative and keep tuning that balance.