OFFICIAL CURRICULUM VITAE (CV)
KEVIN LUO (Luo ZihJia) · 完整履歷規格
KEVIN LUO (Luo ZihJia)
AI Product Manager | Full-Stack AI Engineer | Enterprise AI Consultant
PROFESSIONAL SUMMARY (專業背景與產品實績)
Background & Track Record
- Cross-domain AI Product Manager & Full-Stack Engineer with 5+ years spanning product management, AI engineering, medical AI, and semiconductor manufacturing.
- Transitioned from nuclear physics to AI — a technical-background PM who speaks both the language of engineers and business stakeholders.
- Built 3 complete AI products from 0→1 as sole developer & product owner: LegalSign AI (LangGraph + CopilotKit), FormosaEagle AI (Google ADK + Flutter + VIVE Eagle), and Casper AI (React + Gemini multimodal).
- Contributed as core team member on 2 enterprise-grade platforms: Medical VLM Pipeline (RL + multimodal medical imaging) and ESG RAG System (7-agent deep research platform).
- Independently managed full business cycle: client negotiation → design → development → delivery → payment closure.
PM Leadership & Recognition
- UC Berkeley AgentX Program (Jul 2025, online) — team resource allocation, cross-functional coordination, placing right people in right roles.
- Reinforced PM leadership skills on top of strong technical foundation.
- Paid AI consulting: Deloitte (automation process), Nike & Hotai Motor (Legal AI & product advisory).
- Two-time national AI competition champion | 2 papers at UC Berkeley AgentX.
- Taiwan AI Academy Executive Program graduate (Class 9, 300+ executives).
KEY STRENGTHS (六大核心優勢)
Client-Facing & Revenue
Consulted for Nike, Hotai Motor, Deloitte. Independently negotiated, delivered, and closed payments.
0→1 & Team Builder
3 AI products solo + 2 enterprise platforms as core team. Ships and monetizes.
Enterprise AI Speaker
Trained at Deloitte, HSIP, top universities. Bridges AI and business for executives.
Technical PM Leadership
Technical background = engineer credibility. AgentX = resource allocation & leadership.
Cross-Functional Coordination
Bridge between product, engineering, design, business. Ambiguity to delivery.
Full-Stack Architecture
React/Flutter, FastAPI, Vector DB, LLM orchestration, Docker/AWS deployment.
ENTERPRISE CONSULTING (企業顧問與演講)
Jan – Jun 2025
Deloitte Taiwan
AI Automation Process Consultant
- Independently secured consulting engagement
- Advised on integrating AI-driven workflow automation into enterprise operations
- Designed corporate training on Enterprise AI Applications and No-code Agent development
- Provided AI tool selection strategy and implementation roadmaps
2025
Nike
Legal AI Consultant & Product Advisor
- Advised on AI-augmented contract review, compliance checking, and legal document management
- Assessed existing legal operations and proposed enterprise-scale AI solutions
2025
Hotai Motor (和泰汽車)
Legal AI Consultant & Product Advisor
- Evaluated AI-powered contract review for automotive procurement and dealer agreements
- Proposed phased AI adoption strategy with system integration plan
TECHNICAL STACK BREAKDOWN
AI / ML:
LangGraph, RAG, Multi-Agent, Google ADK, A2A, MCP, Unsloth, TRL (SFT/GRPO), VLM Fine-tuning, LoRA/QLoRA, RLHF, DPO, GRPO, Gemini 2.5/3, OpenAI, XGBoost, CV, Context Engineering
Backend & Cloud:
Python, FastAPI, Node.js, LangChain, LlamaIndex, WebSocket, WebRTC, GCP (Cloud Run, Vertex AI), AWS (App Runner, ECR), Docker, CI/CD
Frontend:
React 18, Next.js 16, Flutter, CopilotKit, Tailwind CSS, Framer Motion, Vite
Databases & Automation:
Qdrant (Vector DB), PostgreSQL, MongoDB, Redis, SQLAlchemy, SQLite | n8n, MCP, CI/CD, Firecrawl, OpenTelemetry
PRODUCT & TECHNICAL EXPERIENCE (經歷與系統規格案例)
LegalSign AI (律果科技)
Technical Lead & Product Owner
Mar 2025 – Present · Taipei
- Identified market gap: Taiwan’s 1.67M SMEs (95% without legal staff)
- Validated through user research, customer interviews, and competitive analysis
- Architected entire platform solo: Next.js 16 + React 19 + CopilotKit UI, FastAPI + LangGraph, Qdrant + MongoDB + MySQL
- Designed 5+ specialized LangGraph workflow agents: Legal consultation (RAG), Contract review (risk scoring), Contract drafting, comparison, translation
- Engineered Playbook version control — competitive moat for client retention
- Deployed on GCP Cloud Run with CI/CD, 99.5% uptime
▶ SYSTEM CASE STUDY: AI legal assistant SaaS with 5 LangGraph workflow agents serving Taiwan’s SME market
1. How does it work?
- User uploads contracts or asks legal questions
- Supervisor agent routes to specialized sub-agents
- Each agent executes LangGraph workflow: RAG retrieval → analysis → response generation
- Results aggregated with risk scores and actionable recommendations
2. Problem Solved:
95% of Taiwan’s 1.67M SMEs lack dedicated legal staff. Traditional consultation costs NT$5,000–50,000/session. Platform reduces cost by 90%+ and delivers results in minutes.
3. Key Mechanisms:
- LangGraph state machine for multi-step legal reasoning
- Adaptive RAG with Qdrant vector search for case law retrieval
- Playbook version control (Git-like) for enterprise legal rules
- CopilotKit for real-time human-AI collaboration UI
Next.jsReactCopilotKitFastAPILangGraphLangChainGemini 3OpenAIQdrantMongoDBGCPDockerCI/CD
Medical VLM — Full Training Pipeline
AI Engineer (Team Collaboration)
Jun 2025 – Present · Taipei (LegalSign AI)
- Collaborated with team to build complete multimodal medical VLM fine-tuning pipeline
- Pipeline stages: Data preparation (MIMIC-CXR, PathVQA, DermaBench) → SFT supervised fine-tuning (Unsloth + TRL) → GRPO RL training → LoRA adapter merging
- Implemented Unsloth framework: 2x training speed, 70% memory reduction
- Designed GRPO reward functions: format compliance + clinical accuracy + CoT quality scoring
▶ SYSTEM CASE STUDY: End-to-end medical VLM training pipeline for multimodal medical image understanding (Radiology, Pathology, Dermatology)
1. How does it work?
- Stage 1: Curate medical datasets into standardized conversation format
- Stage 2: SFT fine-tuning on base VLM (Qwen2.5-VL / Llama 3.2 Vision / Gemma 3)
- Stage 3: GRPO RL training with custom reward functions for medical reasoning
- Stage 4: Merge LoRA adapters → deploy production model
2. Problem Solved:
Generic VLMs hallucinate on medical images. Fine-tuned models achieve clinical-grade accuracy for specific modalities, enabling reliable automated reporting.
3. Key Mechanisms:
- Unsloth weight-sharing: 90% VRAM reduction for RL training
- LoRA/QLoRA selective layer fine-tuning (vision vs. language layers)
- GRPO: no value model needed, more efficient than PPO
- Custom reward scoring: format + medical accuracy + Chain-of-Thought quality
UnslothTRLPyTorchLoRA/QLoRAVision TransformerRLHFDPOGRPOHugging Face
ESG RAG System (LaplaceAI Co., Ltd.)
AI Engineer (Team Collaboration)
Jun 2025 – Present · Taipei
- Built enterprise-grade ESG report generation platform
- Architecture: LangGraph Supervisor-Orchestrator with 7 specialized agents (Research, Content, Compliance, Quality, Review, Finish, ReAct)
- Dynamic multi-expert collaboration: 1–8 experts running in parallel for consensus
- Hybrid RAG: Qdrant vector + BM25 keyword + multi-strategy reranker
- Integrated 8 LLM providers with automatic fallback chain
- Deployed via GitHub Actions → Docker → AWS ECR → AWS App Runner
▶ SYSTEM CASE STUDY: Enterprise-grade ESG reporting AI platform with 7 specialized agents + deep research + hybrid RAG
1. How does it work?
- User initiates ESG report request
- Supervisor routes to specialized agents
- Research agent deep search (Tavily/SerpAPI/DuckDuckGo)
- Content agent drafts | Compliance agent checks regulations
- Quality & Review agents score and refine | Finish agent assembles final report
2. Problem Solved:
ESG reporting is mandatory but extremely labor-intensive (weeks of work). Platform automates research, drafting, and compliance — reducing production from weeks to hours.
3. Key Mechanisms:
- LangGraph Supervisor-Orchestrator for complex multi-step workflows
- A2A (Agent-to-Agent) protocol for inter-agent communication
- Hybrid retrieval: Qdrant semantic + BM25 keyword + reranker
- PostgreSQL checkpoint for crash recovery & resume
- Multi-LLM factory: 8 providers with automatic fallback
FastAPILangGraphLangChainLlamaIndexQdrantPostgreSQLMongoDB AtlasDockerAWSOpenTelemetryA2A
UC Berkeley AgentX Program
Researcher & Team Coordinator
Jul 2025 · Remote (US-based)
- Intensive online research and competition program on cutting-edge AI agent systems
- Team resource allocation and cross-functional coordination; placing right people in right roles
- Published 2 research papers on Edge AI and DB-RAG agent systems
- Validated LegalSign AI product concept in international market
FormosaEagle AI × HTC VIVE Eagle
Product Planning & System Architect
Nov 2025 – Present · Taipei
- Designed 5 parallel AI Agent system using Google ADK Supervisor + Sub-Agent orchestration (Flight, Hotel, Food, Transport, Experience)
- Developed 18 API endpoints (FastAPI) + 12-screen Flutter app
- Real-time voice: OpenAI Realtime API + WebRTC
- Built 10,000+ tourism data entries in Qdrant vector DB
- Advanced to VIVE Eagle Developer Program semifinals as sole individual participant
FlutterFastAPIGoogle ADKA2AMCPOpenAI Realtime APIWebRTCQdrantMongoDBRedisGemini 2.5GCP
National Taiwan University — Insight AI Center
AI Engineer & Researcher
Jul 2024 – Feb 2025 · Taipei
- Won national AI championship (Taoyuan Business Innovation) — Multi-Agent + RAG system POC
- Delivered multiple industry POCs: Automated market analysis reports, Industry knowledge retrieval, Elderly anti-fraud AI
- Led VLM + DPO reinforcement learning study group; Published 2 papers at UC Berkeley AgentX
LangChainLangGraphMulti-AgentRAGGPT-4GeminiXGBoostStreamlitReactFastAPIMongoDBQdrantGCP
FunShine Group (Marketech International)
Medical AI Engineer
Jan 2023 – Jul 2024 · Taipei
- Built mammogram tumor detection (Vision Transformer, clinical-grade accuracy)
- Developed liver image segmentation for surgical planning
- Collaborated with Mackay Memorial Hospital on ML project
- Enhanced models through RLHF, DPO, GRPO reinforcement techniques
PyTorchVision Transformer (ViT)Medical ImagingCNNRLHFDPOGRPOPython
Earlier Roles (ASE Group, Shuttle Inc., NSRRC)
Computer Vision Engineer / AI PM / Production Line Manager
2018 – 2023 · Taiwan
- ASE Group (日月光) — Computer Vision Engineer: Deployed YOLO/CNN defect detection on semiconductor production lines
- Shuttle Inc. (浩鑫) — AI Product Manager: Managed AI product lifecycle for enterprise clients; Taiwan AI Academy Class 9
- NSRRC (國家同步輻射研究中心) — Production Line Manager: Managed synchrotron radiation facility; built systematic analytical thinking
AWARDS & RECOGNITION (競賽榮譽)
CHAMPIONTaoyuan Business Innovation & Entrepreneurship — Multi-Agent + RAG System (2024)
AWARDTaipei Smart City Hackathon — AI Multi-Agent Anti-Fraud for Elderly (Corp. Support + National Finals, 2024)
FINALISTTourism Bureau National Hackathon — Multi-Agent Smart Tourism + VIVE Eagle (2025)
FINALISTPrudential Insurance Hackathon — XGBoost + LLM Prediction System (2024)
PUBLISHEDUC Berkeley AgentX — 2 Papers on Edge AI & Medical Multi-Agent Systems (2025)
EDUCATION & CERTIFICATION (學歷與證照)
- Taiwan AI Academy (台灣人工智慧學校) — Executive Program (Class 9): ML, DL, AI business strategy with 300+ executives (2021)
- UC Berkeley AgentX Program: Online research & competition, 2 published papers (2025)
- Languages: Mandarin (Native), English (TOEIC 720)