Part-time Contractor at Snorkel AI (2025-09 – Present)
- Contributing to Terminus Bench, an open-source coding agent evaluation benchmark providing sandboxed environments for safely executing and testing AI-generated code.
- Designed and implemented new coding agent eval benchmarks to expand task coverage and improve assessment reliability across diverse agent frameworks.
- Fine-tuned DeepSeek V4 Flash using LoRA SFT on curated domain-specific instruction datasets; implemented the end-to-end training pipeline with ms-swift framework, managed data preprocessing and tokenisation, and integrated the fine-tuned model into the benchmark evaluation pipeline to assess performance gains on coding agent tasks.
- CyberSWE Security Benchmarking Project: Developed benchmark tasks to evaluate AI coding agents on realistic vulnerability remediation. Seeded application-layer flaws in real repositories, wrote symptom-only task prompts, implemented hidden blackbox security tests, built deterministic Docker environments, and validated gold fixes through Harbor/Realm oracle workflows.
Senior LLM Engineer at Weibo (2024-10 – 2026-01)
- Built and optimized LLM inference deployments using vLLM on NVIDIA and AMD GPU environments, and worked with vLLM-Ascend for heterogeneous accelerator serving.
- Deployed multiple quantized LLMs and tuned serving configurations for GPU memory efficiency, throughput, and low-latency production inference.
- Leveraged vLLM prefix caching for shared prompts and reusable context, reducing redundant KV-cache computation and lowering time-to-first-token and end-to-end inference latency.
- Implemented a multi-agent report generation system like DeepResearch, enabling multi-turn retrieval and report generation.
- Using RAG as the knowledge source to build an Elon Musk-like AI agent with both short-term and long-term memory capabilities with Qwen model.
- Led the backend architecture for a multi-turn dialogue system. For the Follow-up Question module, built a user profiling-based precision recommendation algorithm. Leveraged high-quality conversational data for supervised fine-tuning (SFT) and applied reinforcement learning (RL)-based optimisation (DPO), resulting in significant improvements in dialogue quality and substantial increases in transition rate and click-through rate (CTR).
- Applied a Search-R1-inspired RL framework to optimize a custom search system, enabling the model to dynamically generate search queries and iteratively refine answers; integrated retrieved results into reasoning loops to improve accuracy and relevance of responses in complex, multi-step queries.
- Designed structured reasoning outputs to guide the model's search and decision-making process, enhancing both precision and interpretability.
- Implemented a modular pipeline supporting multiple LLMs and search backends, achieving significant improvement in retrieval efficiency and answer quality over baseline prompt-only approaches.
Software Engineer at Microsoft (2022-12 – 2024-10)
- Implemented llama3, GPT-4o, Mistral and other base model fine-tuning pipeline on Azure ML studio. Managed training data on Azure blob container. Generated pipeline SBOM files (data tokenised, preprocessing, verification, prompt instructions, LoRA fine-tuning, deployment, evaluation).
- Implemented English and 10 multi-language (Chinese, Korean, etc.) training data pipelines for GPT-3.5 turbo, GPT-4 text summarization and fine-tuning evaluation.
- Used GPT-4 for data augmentation on small datasets to fine-tune GPT-3.5 turbo, LLaMA 3 models in English/multi-language. Evaluated fine-tuned models on Microsoft Teams meeting summaries.
Software Engineer at Kafang Tech (2019-09 – 2021-05)
- Designed and implemented ML training cloud-native platform. Deployed across two Kubernetes clusters. Built user authentication.
- Integrated Redis caching, reducing data retrieval time by 60% and improving app performance.
- Built 10+ RESTful APIs for training data management (upload, validation).
- Implemented Redis task queue to handle training tasks and decouple sidecar/gateway; built gRPC connection between gateway and queue.
- Server monitoring and alerting: Implemented dynamic memory alert rules in Prometheus (Golang, Python). Reduced false alerts from 46.6% to 17.6%. Configured email/Slack alerts in Alertmanager; visualized metrics with Grafana.
- Mail notification: Used go-mall package to create email notification module for user account management.
