Projects

Skyworks Solutions, Inc.

Co-op, Digital Supply Chain Team  |  06/2026 – 12/2026 (Expected)

• Architected a modular agentic workflow in Python that pairs deterministic SQL parsing (sqlglot, pandas) with LLM-based analysis to map SAP Material Master fields onto supply-chain bot logic, giving the team full data lineage visibility.
• Structured the pipeline as node-based orchestration so each parsing, mapping, and review stage stayed independently replaceable, and built it with AI-assisted development (Cursor).
• Implemented a human-in-the-loop review layer, persisting verified mappings as golden references to accelerate root-cause analysis and bug isolation across automated bots.

• Refactored rolling-horizon linear programming (LP) optimization pipelines across SAW and BAW technologies by embedding custom Python modules into KNIME workflows.
• Reduced optimization runtime by 89% (7.5 min → 50 sec) and streamlined workflow design to decrease node count by 70% while preserving identical planning logic.

• Formulated a mixed-integer linear programming (MILP) model using Python (Pyomo/PuLP) and HiGHS to disaggregate weekly production plans into lot-level daily schedules.
• Modeled tool capacity, kit constraints, program constraints, part-tool mapping, material supply constraints, and whole-lot integrity to ensure feasible, executable daily schedules.
• Built a high-performance pipeline to solve within 15 seconds, cutting RF semiconductor de-commit rate below 1%.

Lehigh University, Department of Industrial and Systems Engineering

Ph.D. Candidate  |  08/2022 – Present

• Proposed a novel bilevel optimization formulation for server-side defense in one-shot collaborative machine learning, learning client-level aggregation weights to mitigate biased proxy data from unreliable clients while enforcing fairness constraints.
• Developed the framework in Python and PyTorch with a MySQL store for experiment results and client data partitions, achieving improved fairness with minimal accuracy loss even when unreliable clients make up the majority of the system.

• Formulated a bilevel optimization model for distributed learning, integrating data distillation and unfairness mitigation to generate representative synthetic datasets, ensuring fairness, robustness, and privacy in server-side training.
• Implemented a one-shot machine learning framework in Python and PyTorch with synthetic data generation and differential privacy, reducing client-server communication to enable scalable collaborative learning.

Advanced Micro Devices, Inc. (AMD)

Software System Designer, AsiaOps Planning (TW Team)  |  04/2021 – 07/2022

• Delivered an end-to-end Excel VSTO add-in in C#/.NET over a SQL Server backend for long-range wafer capacity planning across business units and process nodes, single-handedly covering specification, schema design, implementation, and ClickOnce deployment.
• Automated the merge of five worksheets from a Kinaxis RapidResponse export into a single standardized planning sheet with live inventory and days-of-inventory formulas across six rolling quarters, locating each column block by header text at runtime so the tool kept working when column positions shifted upstream.
• Designed a versioned database schema that locks each finalized planning cycle as immutable, and built configuration-driven delta reports between any two versions, replacing scattered spreadsheet copies with an auditable planning record.

• Automated the weekly supply-plan reconciliation behind an Available-to-Promise (ATP) planning portal, shifting the production-control team from computing adjustments by hand each week to reviewing and releasing a system-generated plan.
• Encoded the netting rules as an engine computing gaps between factory build commitments and planned supply at part, warehouse, location, and origin level, then redistributing shortfalls and surpluses across weeks to hold quarterly totals intact, cascading across warehouses and interchangeable parts when a single week cannot absorb them.
• Built the SQL Server backend across five upstream sources, supporting two product groups whose supply data differed in both source and granularity, and authored the functional specification aligning planning, production-control, and business-unit stakeholders in Asia and North America while a front-end engineer built the web UI.

AU Optronics Corp. (AUO)

Senior Software Engineer, Production Planning Information Team  |  09/2017 – 04/2021

• Reformulated the material allocation model into a lightweight optimization model for critical-material pegging, enabling sales teams to confirm order feasibility in near real time during the COVID-driven surge in panel demand.
• Developed the solution end-to-end, including model formulation, SQL-based data preparation, and a C# optimization engine; partnered with an external contractor on the VSTO front end.
• Reduced solve time to minutes for near-real-time buyer decisions, cutting buyer workload by 95% and doubling order success rate.

• Developed a smaller-scale scheduling model based on the LCM DPS framework, incorporating multi-stage process linking across two to three sequential production stages.
• Led and mentored a 3-member engineering team on MIP modeling fundamentals.
• Replaced rule-based scheduling with the optimization model, cutting operational runtime by 90%.

• Delivered AUO's first fully in-house MIP model, optimizing 8-week WPS material allocation across products to minimize de-commit rate.
• Modeled production priority, BU/BG allocation priority, cross-grade sequencing, and BOM constraints, including substitute materials and unit limits.
• Built the pipeline end-to-end: C#/ASP.NET optimization engine, automated SQL Server nightly job, and a VSTO dashboard for review and adjustment.
• Generated production-ready allocation plans each morning based on the prior day's submissions, balancing a ~2-hour optimization runtime with operational requirements.

• Pioneered AUO's first MIP-based production scheduling project, translating 4-month MPS inputs into site/line/shift/jig-level production decisions.
• Modeled part-line, part-jig, and cell-part mappings, grade priorities, shift and cell supply constraints, and jig/line capacities to minimize throughput de-commit rate.
• Led model formulation and validation, partnering with an external solver vendor and using GMPL/GLPK prototypes to verify formulations and diagnose solution issues.
• Integrated a heuristic layer with the MIP model, reducing solve time while maintaining near-optimal solution quality for large-scale binary-variable instances.