Wen-Hsin Hsiao

Role: Project technical point of contact｜Period: Aug 2024 – Apr 2025

An on-premise RAG system built under strict security constraints. The client required that data never touch external cloud AI services, so the entire system uses locally deployable open-source components and runs without any external API.

System Architecture

• Document processing: parses PDF／Word／Excel process and quality-inspection documents into a unified indexable text format
• Retrieval: LlamaIndex pipeline with ChromaDB for vector storage, plus a Reranker for two-stage reordering to improve context precision
• Generation: gemma3:27b-it-qat deployed on-premise via Ollama, using the quantized build to fit local hardware
• Data layer: PostgreSQL／SQLAlchemy records dialogue history, retrieved source passages, user feedback and system logs, so every response can be traced back to its source document
• Interface: Gradio combines Q&A and knowledge base management, wrapping file upload/deletion and index rebuilding into a graphical workflow
• Operations: hardware resource monitoring and system logging modules

Quality Validation

Adopted RAGAS to quantify system performance across answer relevancy, faithfulness, answer similarity and context precision, giving the client and the team a shared language for evaluating results.

Full project title: “Digital Support, Barrier-Free Communication: Development, Support and Promotion of AI-Assisted Communication Aids for People with Speech and Language Disabilities”

One RAG architecture with eight independent knowledge bases, helping people with speech disabilities and their caregivers look up assistive-device applications and event donation information across seven organizations.

System Architecture

• Core: RAG vector retrieval with a Reranker, reducing hallucination and improving answer accuracy
• Knowledge base: crawled each organization's website and merged it with existing FAQs, keeping a separate knowledge base per organization to avoid cross-contamination
• Generation: LLM API integration handling multimodal input and cross-lingual queries
• Interface: LINE Messaging API, so users do not need to install anything new
• Deployment: 8 independently deployed services on Google Cloud Run within a single GCP project

Replicability

The core architecture stays fixed; differences between organizations live only in the knowledge base content and dialogue settings. Adding an organization requires no new development — copy the deployment, swap the knowledge base, go live. Eight rollouts were completed during the project, and the systems have run for nearly 2 years serving 2,000 cumulative messages.

★ 1st Place, 2026 UST AI Agent Innovation & Entrepreneurship Competition (course project)

• Built a department-specific AI agent to address administrative knowledge that lives mostly in TAs' heads, with regulations scattered across sources that students must hunt down themselves
• Uses an AI agent + tool calling architecture: a GPT decision agent reads user intent, dynamically selects the right tool (calendar lookup, graduation credit rules, oral-defense SOP and 7 tools in total) to fetch data from the database, then has GPT compose a structured response
• Developed in Python and deployed on Zeabur; the modular design means the system can be moved to another department simply by swapping the knowledge base

Role: Technical lead of a three-person international team (course project)

• Worked with two GMBA international students through the whole process — user interviews, business model design and implementation — tackling the problem that international students in Taiwan cannot read Chinese menus
• Built an AI pipeline on the Gemini LLM API: menu photo OCR → Chinese–English translation with pinyin → dish description generation → AI image generation, so users can see a simulated image of a dish before ordering
• Deployed on Google Cloud Platform (Cloud Run)

Role: Full Stack Developer (course project)

• Built a competition team-matching site addressing students' difficulty finding suitable teammates and forming cross-disciplinary teams
• Responsible for the competition browsing, teammate recommendation and invitation modules
• Deployed on AWS: Amplify for frontend hosting and CI/CD, with Cognito (registration, email verification and login), DynamoDB (user and matching data), S3 (résumé and image storage), and Lambda + API Gateway (serverless functions for recommendation notifications and email delivery)

Capabilities

• Document processing: parsing PDF／Word／Excel into unified indexable text and chunking it
• Retrieval: LlamaIndex pipelines with ChromaDB vector storage and Reranker two-stage reordering
• Knowledge base governance: crawling and consolidating sources, separating bases by organization or domain to avoid cross-contamination
• Quality validation: RAGAS evaluation across answer relevancy, faithfulness, answer similarity and context precision
• Traceability: recording retrieved source passages so every response can be traced back to its original document

In Practice

• Replicated one RAG architecture across 8 organizational deployments — swap the knowledge base and go live
• Delivered a fully on-premise private RAG system, free of external APIs, to an overseas electronics manufacturing plant
• NVIDIA DLI – Building RAG Agents with LLMs certification

Capabilities

• Designing LLM decision agents that select tools based on user intent and compose the results
• Defining tool interfaces and parameter schemas, connecting to databases, rule tables and external services
• Multimodal LLM API integration (image input, OCR, image generation)
• Prompt design and structured response output

In Practice

• IMYOURBUDDY: a GPT decision agent dynamically selecting among 7 tools — 1st place at the 2026 UST AI Agent Innovation & Entrepreneurship Competition
• MenuLens: a Gemini LLM API pipeline running OCR → translation → description generation → image generation

Capabilities

• Building RAG services, AI agents and webhook backends in Python
• Document parsing and data preprocessing (PDF／Word／Excel)
• API integration, error handling and logging design
• Rapid graphical interfaces with Gradio
• Version control with Git

Certification

• TQC+ Programming Language Python 3 (Computer Skills Foundation, Apr 2024)

Google Cloud Platform

• Deployed and operated 8 Cloud Run services within a single GCP project
• Handled service configuration, project member access control and billing account setup
• Diagnosed and fixed errors by inspecting logs in Cloud Logging
• Systems have run for nearly 2 years, serving 2,000 cumulative messages

Other Platforms

• AWS: Amplify (frontend hosting and CI/CD), Cognito, DynamoDB, S3, Lambda + API Gateway
• Zeabur: AI agent service deployment

Capabilities

• Deployed gemma3:27b-it-qat on client hardware via Ollama, using the quantized build to match available resources
• Assembled a complete system from open-source components in an environment where external cloud AI services were not permitted
• Designed hardware resource monitoring and system logging modules
• Handed over a system that non-technical client staff can maintain themselves, with no developer involvement

Capabilities

• PostgreSQL／SQLAlchemy: data layer design for dialogue history, retrieved source passages, user feedback and system logs
• Index building and maintenance for vector databases such as ChromaDB
• Document parsing and unstructured data cleaning (PDF／Word／Excel)
• Data consolidation and statistical analysis: survey aggregation, average calculation and score distribution plotting
• Crawling official websites and merging with existing FAQs to build knowledge bases

2026 · 19th TMI Thesis Award｜Taiwan Management Institute

Thesis: From Naive RAG to Agentic Generation: An Agentic RAG-Driven Dialogue System for Enterprise Knowledge Sharing

Entered my master's thesis in the 19th (2026) TMI Thesis Award and received the High Distinction Thesis Award. Of 1,021 theses submitted nationwide, 255 advanced to the final round after written review and oral presentation stages, and 193 received awards in total — 26 High Distinction Thesis Awards, 64 Distinction Thesis Awards and 103 Honorable Mentions.

Jun 2025｜NTCIR-18 FinArg-2: Temporal Inference of Financial Arguments

• Took part in an international shared task on temporal inference over financial text, comparing fine-tuned encoders (BERT, RoBERTa) against prompted decoders (GPT-4o, Mistral, Gemma) on bilingual Chinese–English financial corpora
• Presented the method and results entirely in English at the venue in Tokyo, fielding technical questions from research teams across countries
• Received both the Best Oral Presentation Award and the Best Poster Presentation Award

Project: IMYOURBUDDY, a department-specific AI agent

• An AI agent with tool calling: a GPT decision agent reads user intent, selects among 7 tools to fetch data, then composes a structured response
• Addresses administrative knowledge that lives mostly in TAs' heads, with regulations scattered across sources
• Modular design — swap the knowledge base to move the system to another department

Project: CYBJ – Create Your Better Journey

First place at the 2025 NTPU USR Sustainable Technology Proposal Competition, responding to university social responsibility (USR) and sustainable development themes through a technology proposal.

Project: TRADIQ.AI, an intelligent US tariff lookup system

Uses AI to look up US tariff classifications and rates, cutting the time trade professionals spend searching regulations.

Project: US Tariff Master｜Finalist

Entered a tariff lookup AI application and reached the finals of the 2025 Smart Innovation Awards — Best AI Awards.

16 May 2025｜2025 International Conference on Information Management (National Taipei University)

Topic: Advanced RAG-Enhanced Generative AI Dialogue System for Enterprise Knowledge Sharing (Session: IS/ISS3)

Co-authored and presented with supervising professor Min-Yuh Day. The paper grew out of the industry–academia project on an enterprise internal knowledge Q&A system, sharing lessons from building an on-premise private RAG system.

Score: 855

Put into practice through fully English conference presentations, teaching assistance for three English-medium master's courses, and collaboration in an international project team.

Issuer: Computer Skills Foundation

Certificate No.: 212240500012100｜Issued: Apr 2024 (no expiry)

Professional design talent certification covering practical Python 3 programming skills.

Issuer: NVIDIA Deep Learning Institute｜Issued: Mar 2025

Building RAG Agents with LLMs

• Advanced retrieval-augmented generation techniques
• Deploying and applying large language models
• Enterprise-grade AI solution architecture
• Performance optimization and scaling strategies

Issuer: NVIDIA Deep Learning Institute｜Issued: Nov 2025

Fundamentals of Deep Learning

• Deep learning principles and neural network architectures
• Model training and optimization techniques
• Convolutional neural networks (CNNs) and image processing
• Data preprocessing and augmentation strategies
• Model evaluation and performance metrics

Issuer: NVIDIA Deep Learning Institute｜Issued: May 2026

Generative AI with Diffusion Models

• Diffusion model architecture and theory
• Text-to-image generation techniques
• Stable Diffusion implementation and customization
• Fine-tuning and applying generative models
• Prompt engineering and optimization strategies

Issuer: NVIDIA Deep Learning Institute｜Issued: Apr 2026

Building LLM Applications With Prompt Engineering

• Prompt design principles and optimization strategies
• LLM application architecture and workflow design
• Advanced prompting techniques such as few-shot and chain-of-thought
• Evaluating and iterating on LLM application performance