My core mission at Vesta is to create ML models that prevent chargeback fraud for ecommerce merchants. I provide technical leadership, providing distinct tasks to modeling teams that coalesce to help the team reach a specific company goal. Key contributor to global ML strategy.Build models for legacy platform serving individual clients. We demonstrate use of 10 concurrent ML models, develop novel features that had not been previously incorporated in the legacy platform, and generally achieve model KS in the 60-70 range while being capability-limited by the aging platform.Build models for next-generation cloud platform. Onboarded 3 major clients plus multiple marketplaces of 10-100 smaller merchants. KS ranges from the 70-80s, with Recall often in the 60-70% range. For one client that sends us a population of ~50% fraud, we achieve 97-99% Recall with KS in the mid 80s. Deploy models to prod in weeks.Develop a new scalable modeling paradigm. Method allows for a consistent base model to onboard clients instantaneously, while enabling the addition of customer-specific features based on availability or importance. Create a modular component for Vesta’s latest model-serving platform. The component is written purely in Python and hosts the team’s ML models. It is accessed by an outward-facing API. Our design allows for multiple models, rules, and instant feature upgrades.The team has included Data Scientists in 5 different time zones, with 24-hour global coverage for our clients. Teams hand off across the globe. I recruit and hire in all three global regions at the Data Analyst, Data Scientist (Jr., Sr., and Principal) and Data Science Manager levels.Our team operates in a modified and flexible Scrum setup, with metrics tracked at the story level and combined ritual meetings. I organize tasks with the scrum master at the beginning of each sprint and track progress throughout. Present at biweekly demo days to the entire company and CEO.

My core mission as the Director of DS at Zelle was to lead the team to create consortium models to stop fraud and scams in the US banking system. We created new versions of legacy models and new models when new products arose (such as Zelle p2p funds transfer). Increased the size of the team by 150% during my tenure. Team submitted 130 concept briefs and converted 50% to trade secrets and 25% to patents. I provided technical leadership, providing distinct tasks to modeling teams to attain company directives. Provided guidance to the team on human resource issues, developed skills, and encouraged innovation.Led development of a new suite of modernized ML models across Zelle’s entire product line. Led development of ACH and check fraud model. Oversaw development of new account score model, identity risk model, and new Zelle fraud/scam model. Led development of Verify Deposit Risk and Verify Payment Risk products. Underneath both products was a set of two models called based on available information. These models were the first machine learning (ML) models at the company. With the new capabilities ML offered, the KS increased from the mid-50s to the mid-80s. Utilized existing feature-store for faster time to prod. Led several greenfield model POCs that did not make it to production despite impressive performance and capabilities. The common theme of these models was incredible performance due to the fact that they used data from across domains, including deposit and payment behavior, p2p, identity, account, and balance information. I acted as a key product and engineering partner, working with technology, devops/software engineering, infrastructure/architecture, and MLOps teams to troubleshoot issues, suggest improvements, and collaborate on applied ML initiatives. I transformed ambiguous business requirements into manageable specifications to define, execute, and deliver ML projects. I was responsible for ensuring clear communication of the value proposition.

Created a new management structure for the DS team at Zelle, optimized for collaboration, creativity, and thought diversity. Democratized project leadership, training new leaders with every round of projects, while giving mentorship and technical direction by collaborating with them on tasks assignment and specific modeling workflows.Led the development of a new generation of Model Governance reports based on new ML modeling principles, including building a template for future reports. These 100+ page reports were used internally as well as by 3rd parties for model validation so that the product could be brought to market based on governance bylaws. Oversaw the development of a bias and fairness toolkit. This became important at the exact time we created it, and we included the methods and results in the aforementioned model governance report. The outputs ensured minimal bias by class, race, gender, age, etc.As a part of our model development, I generated and optimized business cases per model utilizing the confusion matrix to calculate financial impact and drive low false positive rates. Each element of the confusion matrix, TP, TN, FP, FN has a different dollar value associated with it and the combination changes by threshold. We can optimize the threshold for maximum revenue by using this information. Created the requirements and input for an automated dashboard for all current projects, implemented by the Business Intelligence team. This dashboard was used at the executive level to track the KPIs of the team and visualize project status on the road to production. Acted as the expert authority on data throughout the organization, including AI/ML education seminars, committee meetings, and technical design discussion, evangelizing ML and technical concepts to varied audiences, including senior management. Functioned as a trusted advisor and data thought leader.

Led the team that created the first AI/ML production model at Zelle, including sampling the data and features, choosing the algorithm and prediction explanation method, validating the model both internally and externally, and creating documentation for the model for the banks as well as internal use. The team provided recommended strategies to the banks based on optimized metrics. The modeling process we developed was used on later models and ensured that those models met Model Risk Management and regulatory guidelines.Led a team of 4 DS to initiate development of the Verify Deposit/Payment and counterfeit check models.Managed analytics projects, including leading my team to use best practices and execute time-bound model POCs end-to-end. I provided project guidance, operational planning, and technical supervision to team members while providing status updates to upper management. Took concepts from ideation to project delivery through all phases of the development lifecycle, including implementation support through commercialization.

Utilized quantitative coding techniques with tools/methods such as Hive/hadoop/hdfs, SQL, Python and distributed computing on structured data to contribute to ML model development. I used statistical programming, data mining, and supervised/unsupervised modeling algorithms, such as tree-based and clustering. Led POCs for predicting check and ACH returns, creating a unique mobile device identifier, and predicting counterfeit checks. This included sampling, feature engineering, feature selection, model choice and optimization, hyperparameter tuning, and documentation. Submitted several IP briefs to legal.Modernized processes by either converting SAS code to Python or developing new Python code.Led team to maximize the informational output versus input effort to show the results of our experiments quickly so they could be brought to market faster if needed. If an idea didn’t work, we also spent less time on it.

Used reverse engineering to present clients with a true-cost estimate of their own manufactured products. Through subsequent negotiation, we saved up to 10 million USD per SKU per year. Presented and syndicated results of PCB costing directly at client sites and through teleconference. Worked with firm Partners as a technical expert.Developed Python estimator for predictive cost modeling of PCBs. The inputs included 40+ component type counts, board area, and thickness. The output included a cost estimate and confidence interval. Later developed a Convolutional Neural Network (CNN) to extract component counts from the PCB image. I then used the estimator to prioritize SKUs for true-cost deep-dives and worked with the partner to socialize the model with the client.

Acted as the lead engineer for data and algorithms, bringing academic prototypes into commercially realizable technologies. Led a cross-functional team to design and implement the prototype NFP-E single-cell transfection system, which allowed DNA single-cell transfection without damaging the cell membrane.Transformed the manual tool to transfect single cells into a tool to transfect autonomously at a rate of 1 cell/sec. Reduced size by 97% and cost by 99%. Reduced the UI complexity by 92%.Developed a real-time physics-based streaming algorithm on time-series data, implemented with C++ and embedded C. Directed testing of the algorithm and hardware on live human-derived HEK-293 cells. Evaluated feedback on the test data and optimized the performance of the algorithm. Collected transfection statistics.

I worked on the Modeling and Simulation team at Northrop Grumman.Created object-oriented missile countermeasure algorithm, as well as data analytics tools for an aircraft flight computer. Developed algorithms to determine missile range and countermeasure parameters based on video data. Carried out integration and test work on the F-15 infrared search-and-track pod, including a camera interface. Developed Keep Out Zone for high energy laser DIRCM system to protect the testing ground and battlefield from unwanted high energy laser impingement. The high-energy system could theoretically defeat any incoming threat.

Simulated nanophotonic devices utilizing a distributed high performance cluster (HPC) system. Performed interdisciplinary research among scientists, engineers, and technicians, across pure and applied sciences.Performed large-scale numerical modeling using finite-difference time-domain (FDTD) algorithms for simulation of nanophotonic devices. I used these simulations to generate terabyte-sized datasets prior to extraction and post-processing. While there, I wrote an FDTD electromagnetic simulation program from scratch using Python. The results were visualized in MATLAB to drive further design.As an experimentalist, I collected and analyzed data from these experiments in the time-series and frequency domains. I used statistical methods to verify the robustness of the output data. As an engineer, I wrote low-level code for experimental automation and raw data collection.

I worked on simulation and modeling of nanomechanical and nanophotonic devices, and the intersection of the two. Created first-principles model of nanomechanical cantilever using Partial Differential Equation which could only be solved numerically. Completed Finite-Difference Time-Domain models and simulation of optical nanoantennae.Worked on iterative design, fabrication/implementation, and characterization of imagers and integrated plasmonic devices in the short-wave and mid-wave infrared (SWIR and MWIR). I also automated optical experiments using electronic instruments for more robust data collection compared to manual operation.I designed, fabricated, and characterized optoelectronic devices for biosensing and communications purposes, and developed and characterized novel IR imagers and utilized a largely uncommon flip-chip bonder to integrate a plane of sensors onto the electronics of the camera.