Haris Riaz Email & Phone Number

Herndon, Virginia, United States

Applied Scientist Intern @ Amazon Bedrock Science Synthetic Data Team

San Francisco, California, United States

- Focused on the development of efficient NLP models for identifying social health indicators from clinical text, aiding in comprehensive patient health assessments.- Built a diverse labeled training dataset of 1 million+ provider notes using heuristic annotations, zero-shot learning with LLMs, human-in-the-loop active learning and multiple weak supervision strategies.- Used this dataset to train an NLP model that achieved over a 90% F1 score in expert evaluations, a state of the art in pinpointing key social determinants of health.- Deployed this model into production where it is currently capable of batch inference on millions of provider notes every month.- Gained hands-on experience with Hive, AzureML, Snorkel Flow/Snorkel AI, John Snow Labs’ NLP and Docker deployment.

Tucson, Arizona, United States

Advisor: Mihai SurdeanuRA for the DARPA Habitus Project

Nust-Seecs

Worked on my bachelor's thesis: "Handwritten Sequence Recognition with Time Series Transformers".Implemented a Time Series Transformer (TST) - a variant of the vanilla transformer architecture for handling multivariate time series data with modifications to the input embedding and positional encoding layers.Curated a dataset of sequences of IMU sensory data collected from a Myo-armband representing user arm movements corresponding to text-written In-Air. Trained an encoder-only transformer to achieve close to SOTA accuracy on In-Air handwritten character/digit level classification tasks.Experimented with an encoder-decoder version of the Time Series Transformer to reconstruct the full handwritten sequence from IMU sensory data.

Geneva, Switzerland

- Trained 3D U-Net regression models to predict fluctuations in space charge distortions inside a type of particle detector known as the Time Projection Chamber (TPC).- Developed a validation strategy for comparing different UNet model configurations in Tensorboard.- Enhanced the accuracy of the standard UNet model by incorporating parallel dilated convolutions andresidual connections, resulting in superior performance, particularly in terms of RMSE, during training oncoarse-grained input maps.

Geneva, Switzerland

- Internship with AliceO2 group supervised by Gian Michele Innocenti.- Worked on the binary classification of rare signal versus background in ultra-relativistic heavy-ion collisions, as part of an open source library used by the ALICE experiment.- Implemented and compared performances of XGBoost, Random Forest & Keras algorithms on large imbalanced datasets.- Implemented Bayesian search to speed up hyperparameter tuning of ML models by 5X, compared to grid search.

Doctor Of Philosophy - Phd, Computer Science

Activities and Societies: Computational Language Understanding (CLU) LabAdvisor: Mihai Surdeanu Research Interest: Compositional Reasoning.

Bachelor Of Science - Bs, Computer Science, 3.75/4

Activities and Societies: ACM Technical Team, NUST Community Services Club (NCSC), SEECS Declamation Club.Shortlisted for Rector's Gold.