Alexander Gude
Senior Data Scientist / Machine Learning Engineer
Statement
Machine Learning Engineer with 7 years of experience in direct, customer-facing engagements and shipping data science and machine learning products. Led teams in adopting engineering best practices for code review and CI/CD pipelines and drove transition to enterprise technologies to save in operation costs, reduce production bugs, and accelerate feature processing.
Experience
- Estimated the impact of instant bank account opening (as opposed to delayed until card shipment) to customer paycheck deposit attach rates. Took the lead in pre-emptively building anti-fraud checkpoints for the process and worked with legal, compliance, and business operations to get the program green-lit by our external banking partners. Instant bank accounts delivered a 90% increase in paycheck deposits by our customers and a 55% increase in dollar inflows from paychecks.
- Developed high-recall metrics for monitoring fraud across all ACH transfers giving high visibility into developing problem areas. Built automated alerts based on these metrics so bad activity did not slip through any gaps.
- Using the fraud metrics, targeted high precision rules and machine learning models to curtail the worst fraud. Lowered fraud volumes by 12x in 18 months.
- Monitored the fraud metrics for tax returns deposited to customer accounts. Briefed the IRS and state tax administrators on Cash App Tax’s anti-fraud program, including my monitoring, alerting, models, and rules. My work gave them enough confidence to not redirect tax refund deposits from our customer’s bank accounts, which were 18% of all inflows in 2022.
- Deployed first ACH categorization model for Cash App Banking, reducing uncategorized transactions by 50% and increasing tracked payroll deposit volume by 30%. The improved income categorization has allowed us to make more profitable loans by sizing loan offers to expected ability to repay.
- Led 8 data scientists in building machine learning models to detect and stop fraud.
- Drove adoption of engineering best practices by the team, including implementation of peer review for code changes and automated correctness checking, building of CI/CD pipelines for code and model deployment, and added metrics around test coverage and code health, reducing number of P0 production bugs from 2 in the first year to 0.
- Deployed the first in-product, real-time account takeover prevention model at Intuit—launched in production in TurboTax and alerted security to a possible breach within the first week of running. Back-testing showed it would have detected 95% of last year’s stolen tax return downloads.
- Drove migration of machine learning models from Intuit’s on-prem data center to AWS platform without interruption of services, saving $1.5M per year in operation costs.
- Improved the TurboTax Online account takeover model leading to a 90% reduction in wrongly challenged users, stopping 10X as many fraudsters, and shortening feature processing time from 2 hours to under a second.
- Led a team of 3 engineers in investigating the latest computer vision techniques for vehicle re-identification using deep learning and develop a system within 6 months that enabled clients to automatically detect the same vehicle across multiple videos from security cameras. Handed over the new system to customer’s internal development team and provided training.
- Worked as part of a team of 3 scientists to develop an embedding technique to train a convolutional neural network on unlabeled, open-source image data. Built a system using TensorFlow that learned to embed images and text into a joint vector space, allowing customers to perform content-based image retrieval on a corpus of 100M untagged images.
- Designed and implemented a recommender system evaluation framework in Python and Spark and leveraged it to develop a Python-snippet recommender using word embeddings.
Skills
Languages
Python, Scala, SQL, shell script, C++, LaTeX
Tools
Sagemaker, NumPy, SciPy, Matplotlib, Tensorflow, Pandas, Spark, git, Linux, vim