I am a Member of Technical Staff at Microsoft AI, based in Redmond, WA, building multimodal and agentic AI systems that people can actually use. Representative work includes Phi-3 Vision, Florence, CvT, and HRNet.

1. 2025 – present

   Reasoning, agentic, and coding model training.

2. 2024 – 2025

   Multimodal Llama post-training.

3. 2024

   Led Phi-3 Vision and Phi-3.5 Vision, defining a strong generation of compact multimodal LLMs.

4. 2020 – 2023

   Led and co-authored the Florence-1 and Florence-2 projects. Florence-2 selected as a CVPR 2024 oral presentation.

5. 2018 – 2021

   Led and co-authored CvT, HRNet, and SimpleBaseline — three durable reference points in vision research.

1. 05

   2024 · Technical ReportMultimodal LLM

   Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

   Compact models practical on-device, including strong multimodal capability through Phi-3 Vision and Phi-3.5 Vision.

   Paper Model
2. 04

   2024 · CVPROral27M+ downloads

   Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks

   A unified model spanning captioning, OCR, grounding, and segmentation — one architecture, many tasks.

   Paper Model
3. 03

   ICCV 20213000+ citations

   CvT: Introducing Convolutions to Vision Transformers

   An early hybrid CNN-transformer architecture. Among the most cited ICCV 2021 papers.

   Paper Code
4. 02

   CVPR 20197000+ citations

   Deep High-Resolution Representation Learning for Human Pose Estimation

   The HRNet family — high-resolution representations maintained throughout for dense recognition. Among the most cited CVPR 2019 papers.

   Paper Code
5. 01

   ECCV 20182800+ citations

   Simple Baselines for Human Pose Estimation and Tracking

   A high-performing baseline that became a durable reference for pose estimation and tracking research.

   Paper Code

• ★ Florence-2 accepted as a CVPR 2024 oral presentation.
• ★ HRNet — among the most cited CVPR 2019 papers, 7,000+ citations.
• ★ CvT — among the most cited ICCV 2021 papers, 3,000+ citations.
• ★ SimpleBaseline — among the most cited ECCV 2018 papers, 2,800+ citations.
• ★ 1st place — Look into Person Challenge 2019, Single-Person Pose Estimation Track.
• ★ 2nd place — Object365 Challenge 2019, Full track.
• ★ 1st place — PoseTrack Multi-Person Pose Tracking Challenge 2018.
• ★ 2nd place — COCO Keypoint Detection Challenge 2018.