I'm a Project Controls Engineer at CTCI, Taiwan's largest international EPC contractor and a global top-100 firm in ENR's rankings, where I work on cost and schedule control for large-scale power and infrastructure projects. The job is about keeping complex, multi-year projects honest: planning the work, forecasting where it is heading, and tracking the gap in between. Over the past couple of years I've also designed, built, and deployed AI systems that have reached 70,000+ visits, across computer vision, multimodal understanding, and generative media.

The two sides feed each other. Running real schedules showed me where deterministic plans quietly break down, while building AI taught me how to model uncertainty and prove a method holds up. That overlap is where Converge came from, my most recent project and a Monte Carlo schedule-risk engine. Instead of committing to a single, fragile finish date, it answers a more useful question: how likely is that date, and which paths are really driving the risk?

A Monte Carlo schedule-risk engine that quantifies whether a deadline is actually achievable, and proves its own math is correct

Deterministic CPM produces a single finish date and a false sense of certainty. Converge runs thousands of simulations across a project network to show how likely that date really is, which paths actually drive the risk, and how much contingency a target truly needs. It's the project that fuses my project-controls domain with the engineering behind the AI work below.

• Merge bias, decomposed: Separates two effects most tools conflate: mean shift (from right-skewed durations) versus the true Jensen-inequality gap that appears only when parallel paths converge (e.g. commissioning subsystems merging into Integrated Systems Testing).
• Criticality index ≠ CPM critical path: Ranks activities by how often they land on the realised longest path across simulations, surfacing the near-critical paths that deterministic float hides.
• Correlation as an output, not an input: Implements Hulett's Risk Driver Method, where named risk events hit the activities they physically affect, so correlation emerges from structure instead of a hand-entered coefficient.
• A validation bench, because credibility is the product: 67 pytest tests plus an in-app validation view covering analytic agreement, merge bias checked against SciPy numerical integration, Monte Carlo convergence verified at σ/√N (log-log slope ≈ −0.5), and reproduction of Hulett's published 1996 benchmark (with honestly flagged reconstructions, nothing fabricated).
• Stack: Streamlit · vectorised NumPy · SciPy · Plotly, with a clean engine/UI separation. Flagship example: a data-center commissioning network (MEP → IST) with four parallel L3→L4 paths converging at L5 Integrated Systems Testing.

Newest project, built for rigour rather than traffic.

Links: 🌐 Live App | 🗂️ Repository

A hybrid model for dog breed classification and recommendation

PawMatchAI combines a sophisticated CNN backbone with advanced Transformer layers and a specialized Morphological Feature Extractor to create an innovative hybrid architecture for comprehensive dog breed identification. The project features several key innovations and achievements:

• Advanced Architecture: Hybrid model combining Convolutional Neural Networks with Transformer layers for enhanced feature extraction and classification accuracy
• Custom Feature Engineering: Proprietary Morphological Feature Extractor inspired by expert veterinary observation patterns
• High Performance: Achieved an impressive 88.7% F1-score on breed classification tasks
• Multi-functional Platform: Enables users to identify breeds, compare breed characteristics, and get LLM-powered recommendations for optimal matches based on lifestyle preferences
• Business Intelligence Integration: Comprehensive Tableau dashboard transforms breed classification data into strategic business insights, analyzing market segmentation patterns and delivering analytical breed recommendations
• Recognition: Featured on Hugging Face's "Spaces of the Week" with over 35,000+ visits and 14,000 GPU runs
• Links: 🌐 Try the Demo | 🗂️ Explore the Project | 📊 Business Dashboard

Advanced multi-modal system for deep scene understanding

Vision Scout represents a sophisticated approach to visual intelligence, orchestrating multiple state-of-the-art models to transform complex visual data into comprehensive narratives:

• Multi-model Architecture: Seamlessly integrates YOLO11, CLIP, Places365, and Llama 3.2 for comprehensive scene analysis
• Deep Understanding: Transforms raw visual data into human-readable stories and detailed scene descriptions
• Advanced Processing: Combines object detection, image classification, scene recognition, and natural language generation
• Production-Grade Architecture: Three-layer facade design managing 35,000+ lines of code across 70 specialized classes, showcasing sophisticated system design capabilities for enterprise-scale multimodal AI deployment
• Foundational Research: Tackles the fundamental challenge of true scene understanding beyond simple object detection, addressing spatial relationships, functional purposes, and contextual reasoning that form the foundation for future autonomous systems and AI assistants
• Future Vision: Represents ongoing exploration toward achieving genuine scene comprehension, a critical stepping stone toward more general AI systems capable of perceiving, reasoning about, and interacting with the physical world in human-like ways
• Community Recognition: Featured on Hugging Face's "Spaces of the Week" with significant user engagement
• Performance Metrics: Over 11,000 visits and 4,000+ GPU runs within three months, demonstrating strong user adoption
• Links: 🌐 Try the Demo | 🗂️ Explore the Project

Professional AI-powered image enhancement and video generation platform

VividFlow represents a significant advancement in accessible AI content creation, combining professional-grade image-to-video generation with intelligent background replacement and artistic style transfer in a unified platform. The system addresses critical pain points in digital content production by enabling non-technical users to transform static imagery into dynamic, broadcast-quality videos while simultaneously offering sophisticated background synthesis and artistic rendering capabilities. The project features several key innovations and achievements:

• Triple Creative Workflow: Unified platform integrating image-to-video generation, intelligent background replacement, and artistic style transfer within a streamlined three-tab interface
• Intelligent Motion System: Eight curated template categories designed to prevent common generation artifacts, with support for custom natural language prompts and optional AI-powered prompt enhancement
• Optimized Performance: FP8 quantization and Lightning LoRA distillation enable efficient generation while maintaining professional output quality
• Advanced Background Synthesis: Multi-tier segmentation with twenty-four professionally curated scene templates spanning professional, natural, urban, artistic, and seasonal environments, complemented by AI-powered inpainting for precision artifact removal
• Artistic Style Transfer: Six foundational artistic styles and five balanced blend presets transform photographs into distinctive interpretations ranging from 3D cartoon to classical oil painting, with optional identity preservation technology
• Artifact-Free Compositing: Lab color space blending technology ensures seamless subject-background integration with precise edge handling and manual Touch Up refinement capabilities
• Recognition: Featured on HuggingFace's "Spaces of the Week" with over 27,000 visits

Links: 🌐 Try the Demo | 🗂️ Explore the Project

Comprehensive data science and machine learning portfolio repository

The Learning Record repository chronicles a complete journey through data science fundamentals while tackling real-world business challenges. This collection represents hands-on experience solving complex problems across multiple industries and technical disciplines:

• Financial Analytics: Credit card fraud detection achieving 99% AUC performance using XGBoost and Bayesian optimization
• Credit Risk Assessment: Comprehensive credit score classification models with 85% accuracy utilizing ensemble methods and neural networks
• Natural Language Processing: Advanced MBTI personality prediction through sophisticated text analysis and NLP techniques
• Customer Analytics: E-commerce segmentation analysis using K-means and DBSCAN with optimized silhouette scores
• Time Series Forecasting: Retail sales prediction implementing ARIMA and SARIMAX statistical methodologies
• Signal Processing: Human activity recognition from smartphone sensor data using advanced preprocessing and dimensionality reduction techniques
• Audio Classification: Music genre classification from audio features with sophisticated feature engineering and model optimization

Links: 🗂️ Explore the Repository

I write about deep learning architectures, hybrid modeling, and AI system design—translating complex concepts into actionable engineering insights.

🔹 Machine Learning & AI Builder Hands-on in deep learning, computer vision, NLP, and model deployment, with a focus on building useful, explainable, and well-integrated AI solutions.

🔹 Engineering Execution × AI From running construction schedules to orchestrating multi-model AI systems, I carry the same structured, iterative mindset, whether defining MVPs or architecting segmentation pipelines. I bring ideas from concept to working code: a Monte Carlo schedule-risk engine validated against analytical solutions, multimodal fusion with YOLO + CLIP, morphological feature extractors that enhance CNN reasoning.

🔹 Data Quality Advocate I believe in "Garbage in, garbage out", a lesson that applies as much to training transformers as it does to a poorly defined CPM schedule. Whether it's feature engineering for XGBoost or preprocessing for diffusion models, a well-prepared dataset determines everything downstream.

🔹 Product-Minded Developer I don't just build AI models, I develop solutions that address real user needs. PawMatchAI emerged from understanding breed selection challenges; VisionScout transforms complex visual scenes into comprehensive narratives; VividFlow tackles professional content creation for non-technical users; Converge brings quantitative risk into schedule decisions. I prioritize features based on impact, iterate based on feedback, and measure success through actual adoption. The 70,000+ combined visits across my projects reflect this user-first approach, as does my technical writing on Towards Data Science, which translates complex architectures into actionable insights for practitioners.

I'm drawn to work at the intersection of large-scale program execution and data/AI-driven decision support, in environments where rigour, quantitative reasoning, and the ability to actually ship all matter.

I'm especially glad to connect with:

• Teams in program / project controls, planning, and delivery on complex capital projects (data centers, power, infrastructure) who value quantitative risk and analytics
• People building AI-powered products in computer vision, generative AI, or multimodal systems
• Anyone who cares about both technical depth and turning it into something that ships

Whether it's a potential collaboration, a role, or just comparing notes on schedule risk or AI system design, I'd welcome the conversation.

"Every challenge is a puzzle — it's just waiting for the right combination of algorithms and insight."