I build artificial intelligence systems that live entirely on the edge.
Currently, I'm an engineer at Google India, developing on-device AI for Pixel smartphones. My recent work includes co-designing and optimizing Gemma 4, and future iterations, to run efficiently within the strict hardware constraints of the Google Tensor SoC.
Before joining Google, I was in the SF Bay Area building patented, cutting-edge computer vision pipelines for Verkada.
On-device AI for Pixel, including the Gemma family of models.
Computer vision on the edge for security cameras.
Reinforcement learning for self-driving cars.
Computer vision for transportation.
Where my journey into ML started.
Latest · 2026 · Google DeepMind
Co-author of the Gemma 4 Technical Report. I co-design cutting-edge LLM architectures for the Pixel Tensor SoC, and I'm now working on the next generation of Gemma.
Lead inventor of a parallel on-device LLM evaluation framework, which shards evals across a fleet of phones, cuts 100+ hour runs to about 4 hours and gates firmware and compiler releases.
Matches plates against a text database with 99.6% accuracy using a multimodal triplet loss.
HOV-lane passenger counting deployed on highways in New York, San Francisco and Los Angeles.
Real-time recognition of multiple license plates, running entirely on the camera's vision processor. Each plate is followed as a tracklet through the video stream, and per-frame OCR reads are fused with a confidence-weighted voting table, which makes the system robust to motion blur, poor lighting and stacked or multi-line plates.
Motion-triggered recognition for gates and entry points. Recognition starts as soon as a vehicle moves into view, and each character is resolved by consensus across neighbouring frames, with inconsistent reads rejected.
On-camera tracking of people and vehicles. Detections are linked over time using Kalman-filter motion prediction, and the resulting tracks are indexed in an on-device key-value store that the cloud queries on demand, so analytics never require uploading video.
Attribute search over high-resolution crops of people and vehicles (hyperzooms). An on-camera CNN with an inverted residual architecture and semantic segmentation extracts attributes such as clothing and vehicle color, which are stored on the device and served to the cloud through queries.
▶ Watch on YouTube
End-to-end RL agent that maps camera images and waypoints to speed and steering, beating prior benchmarks on no-crash scenarios.
