Gene Yoffe (PhD)

Predictive Maintenance Project (for a Concrete Ready-Mix Producer/Deliver): Currently leading an ongoing Predictive Maintenance Project for electrical equipment at a rock plant. Utilizing machine learning to predict motor failures, optimize performance, and reduce downtime. Analyzing sensor data (vibration, temperature, voltage, and amperage) and employing statistical modeling to establish normal motor benchmarks, resulting in cost savings and extended motor lifespan.

Optimization Algorithm Design, Implementation, and Deployment (for a Concrete Ready-Mix Producer/Deliver): Conceptualized and coded a multi-step optimization algorithm to enhance truck routing efficiency. This algorithm, developed using techniques from linear programming, convex optimization, and combinatorial optimization, effectively manages various constraints, including truck availability, material availability at different batch plants, and diverse order requirements. The deployed solution significantly reduces both total travel time and discrepancies from scheduled load times, thereby boosting operational efficiency and ensuring improved customer satisfaction

Environment: Python, cvxpy, scipy.optimize, Scikit-learn

Document Image Analysis project (for InterFirst Mortgage)

• Created domain-specific deep neural network image document classifier: 85% accuracy after combining visual and textual features
• Created image processing procedures for automated analysis of customer financial documents to assist clients in making decisions for customer eligibility. This reduced company's manual applicant document processing by 50%.
• Developed new analytical image data processing techniques and assessed their effectiveness and accuracy

Environment: Python, OpenCV, Tensorflow, Keras

Stock trading strategy development (for InterFirst Mortgage)

• In separate project developed proprietary Reddit/Wall Street Bets post-flow-based stock trading strategy. Strategy achieved Sharpe ratio of around 2.7 during months of significant boosting on Wall Street Bets.

Environment: Python, Scikit-learn, XGBoost

• Improved payday model (accuracy increase from 90% to 92%)
• Created ad-hoc payroll classification model
• Built AWS based data pipeline
• Worked on bureau feature selection
• Worked on feature engineering
• Established several fraud filters (which helped to decrease cost per funded by 15%)

Environment: Python, AWS, SageMaker, Scikit-learn, Periscope Data, Redshift, Postgres SQL, MySql

Clients: PepsiCo, Bridgestone Corporation, Grupo Éxito, Guitar Center

• Designed, implemented, and productionalized efficient multithreaded SAS algorithm for computing measure of cannibalization of goods

Incorporating cannibalization sub-model into company's signature product increased overall retail demand prediction accuracy by 0.5-3%

• Worked on number of projects concentrating in forecasting retail demand and establishing price elasticity estimates for CPG/Retail clients.
• Used advanced shrinking/selection multiple regression methods for establishing relevant demand drivers and demand driver elasticities
• Wrote numerous programs to support statistical research
• Used cluster analysis to partition universes of stores into demand-like subsets for further efficient demand modeling

Environment: SAS. SAS Macro Language, SAS HPF, R, Linux, Windows

• Designed and developed market data database and related functionality to serve company needs
• Developed application for reconstructing complete pricing history of stock securities
• Redesigned company’s pricing database by merging multiple separate databases onto one comprehensive central resource
• Developed robust application for creating internal firm-wide security identifiers

Environment: C, C++, C# .NET, SSIS, T-SQL, Windows, Microsoft Windows Server, Linux, Unix, OpenVMS

• Designed and developed massively parallel GPU implementation of matrix QR decomposition

Obtained speedups: ×20-30 times

Environment: NVIDIA TESLA 2050, CUDA framework, Cuda C

• Designed and developed multithreaded Newton’s method and path tracker for solving polynomial systems on shared memory multicore workstation

Obtained speedups: ×0.8×n times on n-core workstation

• Added components into PHCpack – widely accepted open source software for solving systems of polynomial equations
• Made algorithmic improvements for multithreaded path tracker application (used quadratic predictor; incorporated automatic differentiation, optimized input)
• Designed and developed massively parallel implementation of automatic differentiation algorithm for evaluation and differentiation of polynomials in several variables on NVIDIA TESLA 2050 using CUDA framework; Obtained speedups:

×10-20 times for complex double arithmetic,

×25-35 times for complex double-double arithmetic

Environment: multithreading: within Linux using C++, pthreads; GPGPU: on NVIDIA TESLA 2050 using CUDA framework/CUDA C

• Taught multiple undergraduate Math courses as a recitation instructor (teaching assistantship duties). Taught courses included all levels of Calculus, Precalculus, Algebra, Linear Algebra, Differential Equations, etc.

• Created efficient ad-hoc model for extracting elevation profiles from noisy and biased data for new GPS product

It improved elevation prediction accuracy from under 50% to 90%

Environment: R, R-studio

• Created custom multigrid method for solving nonlinear partial differential equations

Environment: C/C++, Linux