Hongshen Zhang

Project

ModelCraft Jan 2024 - Now

Tech Stack: Swift, SwiftUI, SwiftData, MLX

1. Designed a multi-step task orchestration system based on the ReAct architecture, enabling LLMs to autonomously invoke and schedule local tools.
2. Built an Agentic RAG workflow with a local vector database, supporting multi-format data parsing (PDF, audio, images, etc.) to improve retrieval accuracy and reasoning performance.
3. Leveraged the MLX inference framework for efficient local deployment, significantly improving inference performance.
4. Implemented a rolling summarization mechanism to mitigate token overflow, enhancing long-context memory consistency and complex task handling capabilities.

Sentiment Analysis for Financial News Mar 2025 - Mar 2025

Tech Stack: Python, PyTorch, Huggingface Transformers

1. Preprocessed financial news data, performed label encoding, tokenization, and dataset splitting.
2. Supervised fine-tuned a pre-trained BERT model using HuggingFace Transformers, enhanced with LoRA for efficient parameter tuning.
3. Applied CrossEntropy loss with L2 regularization and early stopping to mitigate overfitting and improve generalization.
4. Achieved 86% classification accuracy, outperforming traditional ML baselines.

Rednote-like social app Dec 2024 - Jan 2025

Tech Stack: Spring Boot, Spring Cloud, Nacos, MySQL, Redis, Kafka, Elasticsearch, Minio, Freemarker

1. Implemented scheduled publishing for articles and full-text search functionality.
2. Used Freemarker template engine to render article detail pages, improving page load performance.
3. Performed offline article scoring and leveraged Kafka Streams for real-time score updates.
4. Implemented multi-level rate limiting, including Nginx rate limiting, Gateway rate limiting, and microservice-level rate limiting, to enhance service stability.