San Francisco Bay Area
Startup Company (around 70 employees): Built and led the entire large-model AI department from scratch (team of ~10 algorithm engineers + ~10 engineers/testers/product managers). B2B Focus 1. Super Customer Service (for enterprise clients): - RAG + intent recognition + multi-model workflows + industry-specific knowledge. - Continuously optimized to achieve high customer satisfaction. 2. Industry Knowledge Search Engine: - 1) Knowledge graph construction. - 2) Integration of Elasticsearch (ES) + knowledge graph + RAG. 3. Agentic RAG + Deep Research + ChatBI: - 1) Focus on solving multi-hop reasoning challenges. - 2) Orchestration of multiple agents. - 3) Natural Language to SQL (NL2SQL) solutions. C2C Exploration: Rapidly prototyped and validated projects (reaching ~10k users), including: - Text-to-image generation (SD, Flux+LoRA). - 3D generation. - AI-powered PPT creation. - AI fortune-telling. - Digital human avatars.
hackathon-driven MVP validation across B2B and B2C domains. C2C Products 1. Developed an LLM+RPA-powered social agent platform from scratch, automating personalized user interactions. 2. Created an AI travel assistant integrating knowledge graphs and geospatial data, combining agent logic with IoT hardware for real-time recommendations. B2B Solutions 1. Commercialized RAG-based knowledge base solutions for enterprises, focusing on domain-specific data indexing and retrieval optimization. 2. Specialized in cost-efficient LLM deployment: - Fine-tuning models for vertical industry use cases (e.g., legal, healthcare). - Implementing model quantization for edge devices (TensorRT, ONNX runtime).
Founding Engineer @ SenseTime U.S. Team Drove end-to-end deployment of computer vision products and AI infrastructure (split time between Silicon Valley and Shanghai). Key Deliverables: Built an internal toolchain platform for a ~200-engineer team, empowering model training and inference workflows. Training Focus: - Model optimization. （NAS） - Model pruning and quantization. Inference Focus: Built high-performance heterogeneous inference operator services from scratch, supporting: 1. CPU: x86/ARM architectures. 2. GPU: Qualcomm/MediaTek platforms. 3. DSP: Qualcomm platforms.
Worked at the Augmented Reality Team of DiDi US Team. 1. I played a crucial role in developing data collection platforms that effectively gathered extensive video and sensor data for SLAM and 3D reconstruction purposes. 2. Was responsible for spearheading the development of a cutting-edge AR 3D Navigation prototype system. This system seamlessly integrated map services for navigation, utilized advanced OpenGL-based 3D projection and rendering techniques, and relied on GPS/compass-based localization. One of the biggest challenges I faced during this project was implementing an End2End OpenGL-based rendering pipeline in Android, which required a high level of technical expertise and attention to detail.
1. Upon joining the Android team, I contributed to the development of the Tinder Select V3 feature. 2. I provided assistance with code refactoring tasks, including the conversion of legacy code logic pipelines into RxJava stream pipelines, while working on a large-scale app codebase. 3. As the primary developer, I worked on the Tinder Activity hackathon project and successfully delivered it. :)
(CMU HCI Lab) Leveraged Computer Vision and Deep Neural networks to develop a Wearable Cognitive Assistant System for AED devices. You could find the paper by https://www.cs.cmu.edu/~satya/docdir/CMU-CS-20-102.pdf This work is mentored by Daniel P. Siewiorek, Mahadev Satyanarayanan, Roberta L. Klatzky