Led Google's on-device generative AI initiatives to scale the high-performance GenAI inference framework to hundreds of millions of users.

• Tech lead manager of Google's on-device LLM inference framework, LiteRT-LM(https://github.com/google-ai-edge/LiteRT-LM), enabling the deployment of Gemini Nano to hundreds of millions of devices across multiple platforms, optimizing performance for on-device accelerators (blogpost: https://developers.googleblog.com/en/on-device-genai-in-chrome-chromebook-plus-and-pixel-watch-with-litert-lm/), powering

* Chrome built-in LLM API (https://developer.chrome.com/docs/ai/built-in)

* LLM Inference API for Developers (https://developers.googleblog.com/en/large-language-models-on-device-with-mediapipe-and-tensorflow-lite/)

* On-device GenAI on Chromebook Plus (https://9to5google.com/2024/10/08/recorder-app-chromebook/)

* Gemini Nano on Android (https://ai.google.dev/gemini-api/docs/get-started/android_aicore)

* Image generative in Pixel Studio (https://www.theverge.com/2024/8/13/24219655/google-pixel-studio-ai-image-generation-app)

• Achieved world-record inference speeds for running Stable Diffusion models on-device, pushing the boundaries of mobile AI capabilities. (Paper: https://arxiv.org/abs/2304.11267, Blog: https://ai.googleblog.com/2023/06/speed-is-all-you-need-on-device.html)

ML Engineer working on creating tools to assist in clinical note generation using the audio of provider-patient encounters.

• Technical lead of a team to design and build the

infrastructure for ingesting/hosting data to support

product experiment and research development.

• Developed several models to extract medical concepts

from conversational transcripts and published the results

in top conference.

• Established the model evaluation and monitoring

workflow to connect the model performance with product

needs and streamlined the model deployment process.

• Designed a deep learning model to predict the weekly

opening hours schedule from a single storefront image

that reaches 90% precision production bar.

• Established the system/pipeline to launch the above

model that processed over 1.5M photos and contributed to

275K Google Maps edits by Q1 2018.

Research on image registration for fMRI data

Work in Google Hotel Finder team. Research on helping the Finder to choose the most representative main image for each hotel to show in the search result. The proposed approach is proven to be able to correct more than %80 of the bad selected main images in the dataset. Manage large amount of images in the database (Bigtable/SSTable) through multi-threading and remote procedure call.