•    North Agent Cloud: Tech lead of a 10-person team delivering a full-stack platform for building and running AI agents: scheduling and orchestration, sandbox isolation, egress control, credential custody, access control and audit, and elastic scaling, plus metrics and alerting, object-storage mounts, and end-to-end tracing. It runs in production across several deployments.

•    Decoupling the runtime from the sandbox: ReAct workloads mix long, low-load LLM calls with short, high-load tool execution, so the agent runtime is decoupled from the sandbox and each side scales on its own. Sandboxes start lazily and are reclaimed automatically, so they do not sit idle between tool calls.

•    North Sandbox: A multi-tenant sandbox service compatible with the E2B protocol. Per-template warm pools cut sandbox acquisition to milliseconds, and JuiceFS over object storage serves as the sandbox file system, so data persists in real time and volumes can be remounted elsewhere.

•    North Firewall (built solo, from scratch): Transparent egress interception with iptables, HTTPS inspection and audit through CA injection and MITM, and Vault-backed credential isolation: agents only ever see placeholders, so the credential exposure surface shrinks to a single gateway.

•    Agent Team (built solo): A multi-agent parallel development platform. Agents are split by service and module and work in isolated git worktrees and Kubernetes namespaces; a cloud control plane schedules them across machines and harnesses (built on Claude Code, compatible with Codex), routes messages between them, patrols them every minute, and guards them with custom Stop hooks. It also tracks the PR lifecycle, monitors several clusters, and is reachable from the web, an iOS app, a Feishu bot, and a CLI. 1,400+ commits and about 285k lines, with 5 to 10 agents running on it daily (20+ at peak).

•    Autonomous Agent Architecture & Orchestration: Architected an enterprise-grade agentic system based on Tool-Calling and ReAct paradigms. Orchestrated the end-to-end automation lifecycle—from intent recognition and task planning to execution and validation—ensuring data privacy and high concurrency, contributing to $6M in annual orders.

•    Devi Multi-Agent Collaborative Framework: Engineered a project management agent with MCP toolset to automate project creation (parsing PRDs into deliverables), tracking (analyzing status), and management (dynamic reprioritization). Built a talent matching system to auto-assign tasks based on expertise, automating the full "Create-Schedule-Manage" lifecycle.

•    Self-Healing Workflow: Designed a closed-loop CoT agent for operations. Upon incident triggers, the agent retrieves historical cases, generates fixes, and iteratively self-refines solutions based on error logs while generating verification scripts. Improved solution accuracy to 96% and reduced response time to under 2 minutes.

•    General-Purpose Web Agent: Solved dynamic environment challenges by developing a Selenium-based Web Agent. Combined LLM intent recognition with HTML DOM tree parsing and recursive interaction to manipulate unstructured pages. Introduced Visual Language Model (VLM) verification, achieving a 63% end-to-end resolution rate.

•    Action Space Caching System: Constructed a "Script Alignment + Variable Filling + Multi-path Caching" mechanism. Enabled the generalization of single scripts to multiple incidents and achieved a 20x execution speedup by reusing historical paths.

•    LLM Post-training: Led the Supervised Fine-Tuning (SFT) of vertical Ops LLMs. Utilized Unsloth to accelerate distributed training, achieving a 55% relative improvement in domain-specific technical accuracy compared to GPT-4.

•    Inference Acceleration & Industrialization: Deployed distributed inference services via vLLM. Achieved 200+ tokens/s generation speed and 100+ high concurrency, reducing agent interaction costs by 80% to support high-frequency tool calling.

•    Advanced RAG System: Built a multi-path retrieval system mapping hierarchical data to multiple vector stores. Implemented dimensional expansion, AI summarization, Reranking, and non-linear scoring, boosting the Top-3 hit rate from 65% to 88%.

•    High-Performance RAG Stock Predictor: Deployed an agent-based stock prediction system in Docker. Improved processing speed by 65% and reduced costs by 75% compared to GPT-4, achieving 62% prediction accuracy.

•    OneAPI Aggregation Platform: Developed a unified API gateway with cost/account management, integrating OpenAI, Claude, and Gemini. Enabled universal model switching via a single Base URL, reducing manual configuration time by 85%.

•    Enterprise Sentiment Analysis System: Deployed and benchmarked 20+ LLMs on sentiment tasks. Designed a sentiment analysis pipeline integrating Kafka for real-time processing (News-Analysis-Factor Generation) and reduced latency by 80%.

•    Fine-tuning & Prompt Engineering: Led LLM fine-tuning using P-Tuning v2 and (Q)LoRA. Established a prompt management system, improving F1 score by 20% in sentiment tasks and achieving a backtested Sharpe Ratio of 3.

Program: MA in Computational Social Science

GPA: 3.9

Core courses: Machine Learning, Natural Language Processing, Software Engineering, Data Science, Databases, etc.

Honor: Phoenix Scholarship

Program: BSc in Economics and Finance

GPA: 3.93 (Top 2%)

Core courses: Quantitative Finance, Econometrics, Calculus, Microeconomics, Macroeconomics, Financial Management, Corporate Finance, etc.

Honor: Excellence Academy Award (2021), Excellent Student Scholarship (2020)