Haoyu Zhang
Applied Scientist with 2 years of industry experience and 5 years of doctoral-level research. Expertise in LLM fine-tuning, prompt engineering, and building scalable multimodal ML pipelines.
- Role
- Applied Scientist at Amazon
- Location
- Bellevue, WA, US
- LinkedIn followers
- 500 followers
About Haoyu Zhang
As an Applied Scientist with a background in Theoretical Physics, I am driven to…
Experience
Applied Scientist
Feb 2025 — Present · Seattle, WA, US
Multimodal ML System Design: Built a layered pipeline for IP infringement and fraud detection, using lightweight submodels for early-stage filtering and a large language model at the final stage to handle ambiguous cases.• Multimodal feature engineering: Extracted features from transaction logs, user behavior, text content, and images usingadvanced NLP and computer vision techniques to identify counterfeit listings and suspicious activities.• Data Infrastructure: Developed a data processing package to automatically retrieve data from multiple channels, clean andstandardize inputs, and prepare datasets for downstream prompt optimization and model training.• Data Mining: Tackled labeled data sparsity, especially the lack of negatives, by leveraging high-confidence outputs fromearly-stage filter models and validating them with llm-based labeling. This pipeline enriched training data and improvedmodel robustness and classification performance.• Supervised Fine-Tuning (SFT): Performed supervised fine-tuning to align the base model (Pixtral 12B) with task-specificdata, improving accuracy and robustness.• Prompt Optimization Framework: Developed a comprehensive prompt tuning workflow- Instruction optimization: Invented a novel Parallel Thoughts Prompting (PTP) algorithm that uses llm to optimize theprompt by refining a standardized group of questions and statements- Exemplar optimization: Performed random search over the dataset to identify the most effective set of exemplars- RAG: Constructed a vector database and used top-k similarity retrieval to dynamically select exemplars for each input.• Impact: Developed tailored solutions for general text infringement, general image infringement, and brand-specific infringement. Achieved an overall precision increase from 93% to 96.6% and recall improvement from 48% to 65%.
Education
Fudan University
Bachelor of Science - BS, Physics
Northeastern University
Master's degree, Computer Science
University of California, Berkeley
no degree, Physics
Fudan University
Doctor of Philosophy - PhD, Physics
Find verified contacts for anyone on LinkedIn
Unifers gives sales teams verified emails and direct dials, enriched profiles, and outreach that lands in the inbox.
Free plan included · No credit card required
This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.