Résumé

• Developed and deployed an Automatic Defect Classification system for wire bonders using a ResNet-50 image classifier, reducing defect identification and reject-code assignment time from 3-5 minutes to about 30 seconds.
• Developed operator and admin dashboards for production use, adding automated inference, role-based access, API-key management, rate limiting, model versioning, prediction logging, and model monitoring.
• Implemented a Gold Wire Monitoring System for 400 wire bonders that detects abnormal 99.9%-pure gold wire consumption, possible theft, material loss, and machine-wear signals and reduces machine-level analysis time from 15 minutes to 1 minute.
• Implemented an Airflow ETL pipeline and SPC-based detection using control limits with hysteresis, CUSUM excursion, EWMA max-slope, and intra-lot variance checks.

Technologies: Python, PyTorch, FastAPI, PostgreSQL, Redis, React, TypeScript, DVC, Apache Airflow, Docker, Prometheus, Grafana, Microsoft Entra ID, OpenShift, Kubernetes

• Developed Solution Tester, a no-code UI regression-testing framework for TraceLink's OPUS platform that validates configurable customer pages through reusable BDD scenarios.
• Designed the Playwright-based test engine and custom BDD parser, converting human-readable specifications into executable browser tests and integrating critical platform paths into the nightly regression suite.

Technologies: JavaScript, Playwright, BDD

• Developed ART, an adaptive resampling method that updates training distributions using class-wise validation macro-F1 and empirical class priors.
• Benchmarked ART against 11 class-imbalance methods across five tabular, image, and text datasets using 20 fixed-seed runs per method-dataset pair, improving macro-F1 by 2.64 percentage points on average across the tabular datasets.
• Validated results using Wilcoxon signed-rank tests at p < 0.05 and released ART as a reusable PyTorch package.

Technologies: Python, PyTorch, scikit-learn, NumPy, Pandas, Optuna, pytest

• Built a survival-analysis pipeline using TCGA glioblastoma clinical and somatic-mutation data to study mutation patterns associated with patient outcomes.
• Combined PyTorch autoencoders, K-means clustering, chi-square tests, KEGG pathway enrichment, Cox proportional-hazards modeling, Kaplan-Meier analysis, and log-rank tests.
• Identified six candidate survival-associated genes and built an inference tool that generates individualized survival-probability curves from their mutation status.

Technologies: Python, PyTorch, scikit-learn, GSEApy, SciPy, Pandas, NumPy, Matplotlib, Seaborn

• Built a UCI-compatible C++ chess engine combining NNUE evaluation with alpha-beta search, quiescence search, transposition tables, and advanced pruning, achieving 3,500+ Elo on CCRL 40/15 and Blitz lists.
• Trained and integrated a sparse 768-input NNUE model that improved playing strength by about 500 Elo, optimized inference with AVX2/AVX-512 SIMD and Lazy SMP multithreading, and validated changes through SPRT testing.
• Won the Young Talents CEDR computer-chess tournament with Elixir 3.0 JA.

Technologies: C++ 20, Python, NNUE, SIMD, AVX2, AVX-512, Make