YiuKit Cheung

Building something cool ~

- Developed an NLP-driven ETL pipeline using Python (re, DistilBERT) on Microsoft Fabric to cluster inconsistent procurement items via cosine similarity; resolved item-name discrepancies to establish a master database, enabling high-volume spend visibility and a 10% expected cost reduction. - Architected an automated Medallion pipeline on Microsoft Fabric, leveraging Dataflow Gen2 to ingest, transform, and load fragmented departmental data into a centralized Data Warehouse that automated Power BI report rendering and accelerated Ad-hoc analytics delivery by 50%. - Deployed PCA and LDA models using Python (Scikit-learn) on Microsoft Fabric to quantify latent workload variance across departments; productionized insights via Power BI to human optimize resource allocation, driving a 20% increase in workforce utilization. - Spearheaded monthly cross-functional data reviews across 4 departments; translated Power BI dashboard trends into executive-level strategic insights, identifying operational bottlenecks and driving data-informed decisions that increased company EBITDA by 2%.

- Engineered an automated Medallion ETL pipeline in Python to ingest and process multi-modal data (IoT, AWS S3, APIs, geospatial shapefiles) into an ML-ready warehouse; eliminated manual data wrangling, drastically reducing data preparation time by 50% and enabling advanced Deep Learning capabilities. - Engineered an automated MLOps pipeline using Prefect and Python for a custom PyTorch Transformer model; orchestrated iterative retraining on historical predictions to automate the model lifecycle, eliminating deployment bottlenecks and operationalizing continuous prediction delivery. - Optimized Deep Learning training on 8M+ records with extreme class imbalance (<10% positive) using windowed stratified sampling and a custom focal loss function; heavily penalized false negatives to prioritize rare-event detection, cutting training time by 50% and boosting the F1-score by 10%.

- Developed scalable data transformation workflows using Power Query (M) and VBA to replace manual Excel concatenation; standardized disparate government datasets and successfully processed 30M+ rows, eliminating daily manual data wrangling and reducing preparation time by 35%. - Designed Power BI dashboards mapping demographic trends and a surge in youth driver applications across municipal and remote regions; utilized these data-driven insights to support the policy to digitalize the student driver application process, accelerating executive decision-making by 20%. - Developed 50+ pages of business process maps for operational analysis, supporting cross-functional agreement on workflow improvements and driving a 30% efficiency increase in environmental scanning.

- Engineered a multimodal neural network, integrating BERT and ResNet-108, for efficient execution of multi-label classification tasks. - Orchestrated a data ETL (Extract, Transform, Load) pipeline, encompassing stages of data ingestion, streaming, collection, real-time data analytics. - Developed a RDMS database, utilizing advanced SQL techniques to establish multiple relational structures and achieve optimal data normalization. - Utilized Tableau to generate complex data visualizations, employing advanced calculated fields and interactive actions for effective grouping and visual representation of data sets.

- Utilized R to perform survival analysis, execute tailored-designed statistical comparisons, and apply advanced machine learning techniques with optimization method - Developed simulations of stock market dynamics through the integration of various differential equations and stochastic processes, showcasing a deep understanding of random process modelling.