Portfolio Profile

AI Application Engineer | Enterprise AI Architecture & LLM SystemsPixArt Imaging · ex-AUO · 5+ Yrs Applied AI

Turning enterprise AI requirements into secure, production-ready internal products.

Architecting on-prem LLM workflows, guarded NL2SQL, multimodal document review, and automated data pipelines across IC design, legal, and smart manufacturing.

10+

Production AI Systems

Shipped across IC design, legal, HR & fab workflows

On-Prem

vLLM & Guarded Data

Air-gapped serving, KV cache tuning & scoped NL2SQL

+1.43%

Fab Yield · ~390 hrs/mo

−12.62% defects at AUO & 20+ 1-on-1 staff enablement

Bruce Cheng

Profile Snapshot

Bruce Cheng

Information Engineer · GenAI / LLM

Hsinchu, TW · IC Design & Smart Mfg

Career Track

PixArt ImagingIC Design
2025–Present

Enterprise GenAI & LLM Internal Products

AUO Corp.Smart Mfg
2021–2025

AI Project Leader · Yield & Agent Systems

Core Tech Stack

PythonFastAPIvLLMGuarded NL2SQLRAG & VLMLLMOpsDockerSQL

Engineering Governance

Air-Gapped On-PremRole-Scoped AccessPage-Level Traceability

Engineering Playbook

How I Ship Production AI

Combining trust boundaries, model/tool composition, resilient ETL, and inference optimization into maintainable internal systems.

01

Scope & Trust Boundaries

Define air-gapped constraints, server-side permission scopes, and human review gates before model selection.

Air-Gapped ScopeRole-Scoped AuthHuman-in-the-Loop

02

Model & Tool Orchestration

Compose on-prem vLLM, VLM, OCR, ASR, guarded NL2SQL, and page-cited RAG for verifiable outputs.

vLLM & VLMGuarded NL2SQLPage-Cited RAG

03

Resilient ETL & Internal Web Apps

Normalize encrypted or legacy CP950 data and ship containerized FastAPI backends with AI-assisted Web UIs.

Legacy Data ETLFastAPI + Web UIOffline Docker

04

Inference Tuning & Enablement

Optimize prefix caching (KV cache / TTFT) and turn 20+ 1-on-1 consultations into company-wide guides.

KV Cache / TTFTContext Compression20+ 1-on-1 Guides

Portfolio

Case Studies

Turning ambiguous requirements, constraints, and usage scenarios into internal tools and AI applications teams can actually adopt.

Featured ProjectPixArt

Work-Hour Review & Insights Platform

Guarded NL2SQL · Sensitive Time & Attendance Data · Historical Org Modeling

Built an internal work-hour review platform that converts monthly attendance records into normalized, queryable data and exposes manager review dashboards, exception reports, organization context, and conversational analysis through React and FastAPI.

Pain Point

Internal managers had to manually stitch together fragmented, encrypted, and legacy HR exports to review attendance and work hours, with no retrospective organization tracking or self-service conversational query capabilities.

Solution Design

Delivered a unified analytics platform with automated multi-format ingestion, month-scoped historical hierarchy tracking, and a guarded NL2SQL workflow that generates evidence-backed answers within strict role boundaries.

Role

AI Application Engineer

Workflow Signals

Enterprise AnalyticsGuarded NL2SQLHistorical Data ModelingFastAPI & React
Open project detail
PixArt

Multimodal Technical Document Review

Technical Datasheet Review · Multimodal VLM · Web / CLI / MCP Core

Pain Point

Hardware and firmware teams spent days manually reviewing 50+ page technical datasheets for typographical, register table, and diagram consistency errors, while general tools generated overwhelming false alarms.

Solution Design

Unified datasheet review into a shared multimodal pipeline supporting Web, CLI, and MCP, eliminating formatting false positives and delivering localized section-by-section reports in minutes.

Multimodal Document ReviewTechnical Datasheet LLMMCP IntegrationFastAPI & VLM
PixArt

LLM Gateway & Operations Platform

Model Access & Routing · Usage Metering · Secure LLMOps Reporting

Pain Point

Teams used fragmented model endpoints, credential flows, and log scripts, making model adoption, usage attribution, and GPU capacity difficult to understand without touching production systems.

Solution Design

Unified model access and routing with auditable usage metering and a decoupled operations analytics layer, giving teams and supervisors a reliable view of service health, cost, and capacity.

LLM GatewayModel RoutingUsage MeteringLLMOps Reporting
PixArt

Evidence-Grounded Technical Knowledge

VLM Document Extraction · Human-in-the-Loop Review · Source-Cited RAG

Pain Point

Critical engineering troubleshooting rules and design procedures were buried in hundreds of slide decks and PDFs, making keyword search ineffective and knowledge transfer slow and error-prone.

Solution Design

Delivered a traceable knowledge pipeline where VLM-extracted claims are validated, human-reviewed via Streamlit, and indexed into Chroma, providing exact page-level citations for every retrieved answer.

Evidence-Grounded RAGMultimodal VLMVector Retrieval (Chroma)Streamlit Review UI
PixArt

Enterprise GenAI Enablement & LLM Performance Optimization

LLM Observability · Prefix Caching & KV Reuse · OpenCode & Prompt Best Practices

Pain Point

Users relied on ad-hoc prompts and trial-and-error in AI coding workflows, leading to oversized context windows, repeated prefill overhead, API timeouts, and recurring 1-on-1 support requests.

Solution Design

Combined LLM monitoring insights with 20+ 1-on-1 consultations to publish management-announced bilingual web guides covering prompt engineering, context compression, OpenCode Skills, and prefix-caching verification methods.

LLM Performance OptimizationPrompt & Context EngineeringOpenCode EnablementKnowledge Transfer

Career Changelog

Experience

Iterating on my career like software — each major release moves into a new industry context, evolving AI from isolated analysis tools into everyday internal products.

v3.0 · LatestActive in Production

Information Engineer (Generative AI / LLM Projects)

PixArt Imaging Inc.Aug 2025 - Present
PixArt Imaging Inc. Logo

What's New

Expanded into IC design applications — turning LLMs into everyday internal products across Legal, R&D, and engineering.

Release Notes · Highlights

  • NewDesigned a shared internal GenAI workspace and LLM Gateway (vLLM, OpenAI-compatible APIs, model routing, API key governance, LLMOps telemetry) for coding, document review, and technical analysis.
  • NewDelivered secure internal analytics and document systems spanning natural-language SQL (NL2SQL) restricted to read-only manager-scoped access with audit logs, multimodal IC datasheet review, and page-cited ChromaDB RAG.
  • ImprovedProductionized sensitive legal and patent workflows using OCR, local LLMs, Prompt Engineering (clause/paragraph extraction, staged proofreading, terminology rules), and validation checks — helping R&D and legal/admin teams offload manual proofreading to focus on core work.
  • ImpactAutomated engineering and meeting workflows with reusable Python tools, including robust file-processing pipelines, ASR, speaker-aware segmentation, and structured summarization.
  • ImpactIndependently drove enterprise GenAI enablement, knowledge transfer, and LLM optimization — turning telemetry and 20+ 1-on-1 consultations into company-announced Prompt Engineering, OpenCode, context-compression, and KV-cache / TTFT guides.
PythonFastAPIvLLMLLMOpsPrompt Engineering
v2.0 · LTSPast Role

AI Engineer / Project Leader

AUO Corporation - Advanced Manufacturing CenterAug 2021 - Aug 2025
AUO Corporation - Advanced Manufacturing Center Logo

What's New

Moved from academic research to the factory floor — turning AI models from lab benchmarks into daily tools that improve yield and save hours.

Release Notes · Highlights

  • NewServed as Project Leader (2023–2025), mentoring engineers and interns to deliver cross-factory AI projects from requirement scoping and feasibility assessment to production rollout.
  • ImpactImproved yield by 1.43%, reduced defects by 12.62%, and saved approximately 300 man-hours monthly through cross-factory Golden Path optimization and LIME explainable recommendations.
  • ImprovedEvolved an Azure OpenAI–powered manufacturing assistant from RAG retrieval into a LangGraph tool-using agent workflow with SQL/API tools and process-knowledge reasoning.
  • ImprovedReduced defect-prediction RMSE by 20% and saved approximately 90 man-hours monthly through a PyTorch LSTM preventive-maintenance workflow with a custom trend-consistency loss function.
  • NewRecognized as Smart Manufacturing Level 3 Elite Talent; preventive-maintenance outcomes were approved for an internal defensive patent.
PythonAzure OpenAILangGraphPyTorchSQLDatabricksFastAPI
v1.0 · Initial ReleasePast Role

Research Intern

Academia Sinica - GIS CenterFeb 2019 - Jul 2019
Academia Sinica - GIS Center Logo

What's New

Initial public release — first encounter with messy real-world address datasets, building the foundation for data cleaning and web verification.

Release Notes · Highlights

  • NewBuilt an address normalization workflow and supporting data structures for heterogeneous datasets.
  • ImpactWon the Best Student Paper Award at TGIS 2019.
PythonDjangoPostgreSQLData Quality

Capability View

Capability Map

This is not a disconnected tag wall. It is a map of how I place technical choices inside requirement framing, system composition, product implementation, and domain workflows.

Generative AI & LLM Systems

Build practical, verifiable GenAI workflows around OpenAI/vLLM APIs, retrieval, vector stores, tools, and multimodal inputs.

OpenAI / Azure OpenAI & OpenAI-compatible APIsOn-Prem LLM Serving (vLLM)Prompt & Context EngineeringRAGVector Databases (ChromaDB)Agent Workflows (LangGraph, MCP)Multimodal / VLMOCR / ASROutput Validation Design

AI Architecture, LLMOps & Security

Design secure, reliable enterprise AI systems from feasibility assessment and project leadership to LLMOps governance.

Enterprise AI ArchitectureLLMOps / MLOps DeploymentAI Security & Data Privacy (Offline Deployment, Access Control, Audit Logging)Feasibility & Adoption StrategyTechnical Project LeadershipTechnical Enablement & Knowledge Transfer

Backend & Data Engineering

Implement dependable services and data pipelines that make AI workflows usable in production, paired with AI-assisted Web UI delivery.

PythonFastAPIREST APIsSQLPostgreSQLDockerETL / Data PipelinesDatabricksAI-Assisted Web UI Delivery (React / TypeScript)

Model Optimization & Applied ML

Optimize LLM inference latency and apply custom-objective modeling to operational and manufacturing problems.

LLM Inference Optimization (KV Cache / Prefix Caching / TTFT)PyTorchCustom Loss OptimizationTime-Series Forecasting (LSTM)Explainable AI (LIME)

Background

Education

National Yunlin University of Science and Technology
v0.2 · Beta

National Yunlin University of Science and Technology

Master's Degree

Industrial Engineering and Management

2019 - 2021

Thesis / Research

Improve the Predicted Efficiency of Crowd Flow in Large Areas Based on the R-tree and the SPACE-MDL-LSTM

National Yunlin University of Science and Technology
v0.1 · Alpha

National Yunlin University of Science and Technology

Bachelor's Degree

Industrial Engineering and Management

2015 - 2019

Thesis / Research

Optimization of machine dispatching in flexible production processes

Recognition

Awards & Recognition

2025

Smart Manufacturing Level 3 Elite Talent

AUO Corporation

2020

Best Student Paper Award

Taiwan Geographic Information Society Conference (TGIS)

2019

Best Student Paper Award

Taiwan Geographic Information Society Conference (TGIS)