About me

I work as a Fractional Chief AI Officer (CAIO) — helping PE-backed SaaS and fintech companies that are investing in AI but not shipping products that move revenue.

Before this, I spent 12 years building and shipping AI at DeepMind and in production systems reaching over a billion users — recommender systems, RL agents, on-device GenAI features, safety systems. Over 20 launches. $300M+ in revenue impact. The first team to deploy DeepMind research into consumer products at that scale.

I built and led ML teams across the US and Asia, ran an AI incubator (10 sprints per quarter, idea to prototype), and built the largest ML community in APAC — 500+ engineers across products and countries.

That track record is the foundation of the consulting practice. I know what breaks in production, what directions waste engineering spend, and what it takes to build a team that can run AI without a consultant in the room.

Before AI, I worked in fintech — credit derivatives at Bank of America Merrill Lynch.

2023 Mobile AI

Google Pixel

Recorder Summarization - the first foundational language models on mobile devices (press), awarded one of the 10 tech innovation at Google in 2023. LLM-based weather summarization and diffusion-based weather background generation. Bandit-based recommender for Pixel.

2022 Enterprise AI

Developer Tools

NLP-based bug management tools lowered fix latency and costs. LLM-based root cause analysis tools. LLM-based test scheduling offered a double-digit cost reduction. LLM-based test code generation.

2020 Recommender Systems

YouTube

Bandit algorithms for ad timing brought in the biggest CTR increase in product history. Multiple launches in Ads Safety and Trust & Safety contributing $300M+. Noisy labels techniques leading to a publication: arxiv

2018 E-commerce

Google Shopping

Neural-network based recommendations - 5 launches, each with a single-digit increase to #queries, #clicks, #revenue, #interactions. Bandit algorithms for whole-page layout optimization - a launch resulted in decrease in abandoned queries.