Software Engineer · Backend, AI Agents & native macOS apps
Work Experience
ByteDance
Backend Engineer·China
- Old Lark system retirement & data migrationDeveloped and tested data migration scripts for retiring the old Lark system. Migrated data from 300+ enterprises across 20 tables and 20M rows with zero incidents, cutting maintenance costs by 30%.
- Technical owner, Lark HeadhuntingLed technical planning and solution design, took part in requirement analysis, resolved 10+ production issues and streamlined troubleshooting, with zero incidents through the development cycle.
- Unit testing improvementsRaised coverage of legacy code and kept every incremental change above 70% coverage, lifting overall project coverage by 10%+.
Internship
Amazon
Software Development Engineer Intern·China
- Designed and implemented an efficient RAG workflow with Python, Java and AWS for large-scale document retrieval and intelligent answers.
- Built a conversational interface in React Native, optimised for input/output-constrained devices such as Kindle, with smooth, low-latency interaction.
- Developed core native modules in C++ for networking and memory management, keeping the app fast and stable on resource-constrained devices.
Projects
ModelCraft
Jan 2024 – Present- Built a ReAct-based agent orchestration and tool-calling system that autonomously uses local tools such as search, terminal and Computer Use.
- Implemented inference pipelines for Stable Diffusion, LTX-Video and MusicGen in Swift + MLX for on-device image, video and audio generation.
- Implemented KV cache and rolling summarisation for context management, improving on-device inference and long-running tasks.
- Built an Agentic RAG pipeline with multi-format file parsing and vector retrieval.
SwiftSwiftUISwiftDataMLX
Sentiment Analysis for Financial News
Mar 2025- Preprocessed financial news: label encoding, tokenisation and dataset splitting.
- Fine-tuned a pre-trained BERT model with Hugging Face Transformers, using LoRA for parameter-efficient training.
- Used cross-entropy loss with L2 regularisation and early stopping to curb overfitting and improve generalisation.
- Reached 86% test accuracy, outperforming traditional ML baselines.
PythonPyTorchTransformers
Rednote-like Social App
Dec 2024 – Jan 2025- Implemented scheduled article publishing and full-text search.
- Rendered article detail pages with the Freemarker template engine to speed up page loads.
- Scored articles offline and used Kafka Streams for real-time score updates.
- Added multi-level rate limiting (Nginx, gateway and per-service) to keep services stable.
Spring BootSpring CloudNacosMySQLRedisKafkaElasticsearchMinIOFreemarker
Last updated: 19 Sep 2026