Master of Analytics · UC Berkeley

Qijie (Jerry) Yu

Software engineer building AI agents.

I design multi-agent systems, MCP integrations, and the evaluation pipelines that keep them reliable — built on a solid foundation in backend services, APIs, and cloud infrastructure. I like turning a messy business process into an agent that can actually run it, and measuring whether it works.

Looking for roles building AI agents and AI workflows.

qijie25@berkeley.edu +1 510-542-3685 LinkedIn GitHub
86% → 100% warehouse agent accuracy after prompt redesign
−77% token usage in multi-agent negotiation
96.5% ViT accuracy across 201 food classes
−40% load time on a ~1,000-user hospital backend

Experience

Software Development Intern AI Agents

Adas Eco

San Francisco, CA May 2026 – Aug. 2026
  • Automated warehouse task intake by building MCP clients connecting Google Sheets and Gmail, plus weekly reporting through Notion's MCP server — processing ~10 daily requests and eliminating ~2 hours/week of manual compilation.
  • Evaluated intake validation across 100 test cases and found 86% accuracy, with failures largely from the agent not checking available storage inventory; redesigned the system prompt to require internal tool calls for real-time inventory checks alongside forklift load and shipment constraints, reaching 100% across another 100 cases.
  • Built an evaluation and observability pipeline (DeepEval, Arize Phoenix) across all 4 agents, giving future maintainers live tracing and a 19-test regression suite.
MCP FastAPI Next.js PostgreSQL AWS DeepEval Arize Phoenix

Research Assistant AI / RAG

East China University of Science and Technology

Shanghai, China Jan. – Jun. 2025
  • Developed a Vision Transformer (ViT) for food recognition, achieving 96.5% accuracy across 201 classes; optimized the architecture to ViT-Slim for 2× faster inference with only a 2% accuracy drop.
  • Built a pgvector-based RAG system that retrieved relevant dietary guidelines from health ratings derived from ViT food predictions plus users' dietary goals, and generated personalized LLM recommendations.
PyTorch TensorFlow ViT pgvector RAG

Software Engineer Intern Backend

B-Soft Co., Ltd.

Shanghai, China Jul. – Aug. 2024
  • Built a distributed microservices backend for a ~1,000-user hospital system, using Redis, RabbitMQ, and Kafka to asynchronously handle registration requests at peak traffic; served static assets via Nginx on a high-bandwidth server, cutting load time by 40%.
  • Developed Spring Boot RESTful APIs to retrieve patients' unpaid bills, and corrected SQL aggregation logic that had been duplicating calculations between primary and subcategory charges.
Spring Boot MySQL RabbitMQ Redis Kafka Nginx

Projects

Agora Multi-Agent

Multi-agent marketplace for sponsors and event organizers

agentagora.app ↗
  • Built a LangGraph multi-agent marketplace matching sponsors with event organizers; implemented negotiation between sponsor and organizer agents through a mediator agent, with human-in-the-loop approval before generating contracts.
  • Ran 100 negotiation trials and found deals took ~15 rounds on average, driving high token usage; introduced turn-aware system prompting with remaining-round visibility and a 10-round cap, cutting average convergence to ~7 rounds and token usage by ~77%.
LangGraph Multi-Agent Human-in-the-Loop Vercel

Doberman AI Security

Authorization engine for LLM tool calls

github.com/DobermanCore/Doberman-Core ↗
  • Designed an authorization engine enforcing allow/authenticate/block decisions for LLM tool calls across Claude Code and Codex, combining deterministic security guardrails with anomaly detection through hooks and an MCP proxy.
  • Discovered and remediated a guardrail gap in sensitive environment variable protection that was covered only by post-execution pattern matching; added a pre-execution guardrail, validated with 100% regression test coverage.
Claude Code Codex MCP Proxy Guardrails

Pneumonia Image Classification Computer Vision

ResNet-50 with attention modules for CT-based diagnosis

  • Built a ResNet-50 pneumonia classification model in TensorFlow with PSA and CBAM attention modules on 5,000+ CT images, achieving 96% classification accuracy across four pathological types.
TensorFlow ResNet-50 CBAM PSA

Skills

AI & Agent Systems
LangGraph LangChain Model Context Protocol (MCP) Multi-Agent Systems Tool / Function Calling RAG Prompt Engineering Guardrails LLM Evaluation & Observability PyTorch TensorFlow
Languages
Python Java SQL JavaScript TypeScript
Web & Backend
React Next.js Node.js FastAPI Spring Boot REST APIs WebSocket PostgreSQL
Cloud & Infra
AWS GCP Azure Docker Git CI/CD (GitHub Actions) Redis Kafka RabbitMQ Distributed Systems

Education

University of California, Berkeley

Master of Analytics, IEOR · GPA 3.86/4.0

Berkeley, CA Aug. 2025 – Aug. 2026
  • Relevant Coursework: Python Programming, Analysis and Design of Databases

East China University of Science and Technology

Bachelor of Engineering in Automation · GPA 3.77/4.0

Shanghai, China Sep. 2021 – Jun. 2025
  • Relevant Coursework: Data Structures and Algorithms, Java Programming, Data Mining, Machine Vision