Aaron Reich Email & Phone Number

Working as a Senior ML Engineer contractor with Disney on their DisneyGPT AI Platform & models to increase the alignment of their RAG system with users across various internal use cases.Employed as Lead Data Scientist full time with Tavant Technologies.

Glendale, California, United States

Working as a Senior ML Engineer contractor with Disney on their DisneyGPT AI Platform & models to increase the alignment of their RAG system with users across various internal use cases. Employed as Lead Data Scientist full time with Tavant Technologies.

Remote

Alpine Health is a Health-Tech AI company whose flagship product Alpine Navigator integrates with Epic and helps hospitals discharge patients with clinical and social needs. OSF HealthCare is an investor and customer and Alpine is also backed by High Alpha Innovation.Developed proprietary Clinical LLM, Information Retrieval, and Predictive Risk Modeling technology from ideation to maturity, leading R&D efforts working closely with Alpine Health's CEO, Alpine Health's internal healthcare SME's, OSF HealthCare's Innovation Office, OSF's Case Mgt department, and OSF's IT department.

• Promoted to lead ML/NLP work on Cerner’s HealtheHistory product after presenting to Pionetechs' VP of Engineering the generalization advantages of transitioning from the current Expert-based system to a ML-based system• Significantly increased the accuracy of the outcome evaluation processes of HealtheHistory's health record retrieval automation (RPA) processes by transitioning from a large Expert system to a LLM-based Classifier system, which significantly reduced the number of daily orders resulting in human fallbacks to Cerner Operations.a) Constructed the training set leveraging high quality labels from the rule-based Expert model involving samples from web portals found to have the lowest amount of changes on a regular basis and performed cleaning on the webpage extract data samplesb) Generated synthetic examples leveraging rule-based text level augmentation code for upsampling minority class examplesc) Model selection included experiments with models such as SVM & BoW, Bi-LSTM & GloVe embeddings, BERT, and DistilBERTd) Deployed highest performing classifier to AWS Sagemaker and performed continual training processes throughout model lifecycle• Identified how the manual collection processes of the HTML tags powering the dynamic web automation system could be transitioned to a generalizable deep learning NLP system a) Conducted experiments leveraging a token classification model for the classifying of the HTML attribute values, using a PyTorch Character level LSTM model and Character level RNN model to support this automated HTML tag collectionTools used: Python, PyTorch, Scikit-learn, NumPy, Transformers, AWS Sagemaker

• Worked with Oracle Cerner's product management and operations team, and with Pionetechs' VP of Engineering and Software architect to develop Cerner's HealtheHistory product (an enterprise grade medical record retrieval product serving life insurance companies such as New York Life and USAA, and Morgan & Morgan). The system processed thousands of orders on a daily basis.• Built and maintained key automation process (RPA) application infrastructure and provided technical advisory to Cerner Operations allowing them to scale health record retrieval (RPA) coverage to roughly 6,000 data sources. • Responsible for maintaining the RPA outcome evaluation Expert system’s string matching rules and logic through knowledge engineering methodology • Significantly decreased dependance on Pionetechs' development team for RPA development by developing no-code development tools consisting of extensions to an internal ASP.NET website, web automation NET class library, and a Windows Forms automation dev app. Tools used: VB .NET, SQL Server, iMacros, Python

Atlanta

• Worked as a researcher in Dr. Srijan Kumar's CLAWS Data Science Research Group. Researching the effects of social content-recommendation graph-based models on preference amplification and polarization with misinformation data. Built a novel Twitter dataset by crawling the Twitter API and performed network analysis on the collected graphs.• Selected the GraphRec model for use in the experiments due to identifying it as a state-of-the-art social content-recommendation graph-based model used in production in industry. Developed the code for transforming the data for ingestion by the GraphRec model as well as the negative sampling of the retweet interactions. Altered GraphRec’s implementation to classification as well as incorporated the negative sampling into the loss function. Conducted hyperparameter tuning of the custom GraphRec model. • Gave high-level technical presentation to lab on papers from ACL 2022 conference applicable to the lab’s research.

Atlanta, Georgia, United States

• Worked as a researcher in Dr. Diyi Yang's Social and Language Technologies (SALT) Lab. Lead researcher on the research project resulting in Leveraging Expert Guided Adversarial Augmentation For Improving Generalization in Named Entity Recognition published in the Findings of ACL 2022 which received significant industry interest. • Manually constructed a challenging test set for NER from CoNLL 2003 using an identified hard case and empirically found that state-of-the-art NER systems like BERT drastically drop performance when tested on our challenging adversarial dataset. Implemented the Numpy code for automating the construction of these adversarial examples for use in adversarial training.• From the literature review, identified that mixup would be well suited for the circumstances of the problem due to its robustness to adversarial examples. Made novel updates to Mixup’s interpolation to increase performance and conducted experiments with SOTA adversarial training methods. Adapted a FreeLB based model called TAVAT for the NER problem for use on the experiments.

Boca Raton, Florida, United States

• Worked as a researcher in Dr. Kwangsoo Yang's lab. Involved in the formulation of the Nearest Neighbor Triangulation and Merging algorithm published in Size Constrained k Simple Polygons in the proceedings of the 26th ACM SIGSPATIAL Intl. Conference on Advances in Geographic Information Systems. Tested the algorithm for all theoretical cases of polygonal shapes and was involved in its Java implementation.• Involved in the formulation of the Bipartite Space Shrinking algorithm published in Coverage Constrained Spatial CO-clustering. Involved in the reformulation of the Nearest Neighbor Triangulation and Merging algorithm as well as the novel Nearest Neighbor Point and Merging algorithm which outperforms the original version. Published in Geoinformatica 2020.

Boca Raton

Developed applications as well as database design for the building of a data warehouse to enable automation in the reporting of business metrics to executive management. Tools Used: C#, .NET, Transact-SQL, Microsoft Visual Studio 2015, SQL Server Management Studio 2008

Delray Beach

Was in charge of maintaining and building functionality for the company website, along with uploading content and inventory.

Ft. Lauderdale

Worked on JavaScript and PHP based platforms and frameworks specifically Node.JS, Express.JS, WordPress, and utilized MongoDB - a cross-platform document-oriented database.

Master Of Science In Computer Science Specialization In Machine Learning, On-Campus - Atlanta, Ga, 4.0

Worked as a lead ML researcher in 2 GaTech research labs during my MS CS program for academic credit towards my degree. Advised by Dr..

Bachelor Of Science - Bs, Computer Science, Gpa: 3.5/4.0, Graduated Cum Laude

Activities and Societies: Co-Founder of Upsilon Pi Epsilon International Honor Society for the Computing and Information Disciplines.