Maxwell B. Strome

Email: mbs2241@columbia.edu
Phone: (410) 371 6247
Website: maxstrome.github.io

Education

Master of Science in Computer Science, Machine Learning concentration, December 2022
GPA: 3.83/4.00

Type	Advanced	Comfortable	Familiar
Languages	Python, SQL	HTML/JavaScript, Bash	C++, Scala, Java, C
Frameworks	NumPy, Pandas	TensorFlow, PyTorch, OpenCV, HuggingFace, React
Cloud		AWS, Azure

Senior Data Science engineer at DataSense Software, an automated machine learning startup focusing on time series forecasting of sales in the Corporate Performance Management space.
Helped with the creation, maintenance, and future direction of the auto ML engine from data cleaning to modeling. Primary languages are Python and SQL, and cloud architecture is Microsoft Azure.
Reduced database size by 40% on average and optimized query speed leading to 25% faster run times, both allowing for up to 25,000 target runs, less strain on the database, and a faster user experience.
Added feature explainability engine to the pipeline including prediction explanations, feature impact, and prediction intervals using state of the art methods such as SHAP and conformal prediction intervals, providing insights to the user on the predictions of the models and how those predictions were made.
Implemented production level time series forecasting models ranging from statistical models such as ARIMA, machine learning models like XGBoost, and SOTA deep learning models including transformers.
Optimized memory and process allocation on orchestrated virtual machines, decreasing average runtime.

Employed Java to create a pipeline capable of storing and displaying live network telemetry data, utilizing the graph database Neo4j and various visualization tools.

Wrote TensorFlow model to detect kinship given two photos; specifically used an ensemble of two deep fusion networks with the bases being Senet50 and Resnet50.
Final paper accepted at IEEE International Conference on Automatic Face and Gesture Recognition 2021. End model is an ensemble of three different kinship models including the one above along with models written by OpenAI Codex. Won third place in FG 2021 Recognizing Families in the Wild Challenge.

Employed hierarchal clustering during the pre-indexing phase of a database to generate a dendrogram.
Interrogated this dendrogram with a UCB algorithm to accelerate Select Limit queries that have a user defined function as the predicate.
Final paper was accepted to the SIGMOD 2020 conference.