• Building deep learning pipelines in computer vision to do classification and survival prediction on high-resolution histology images that are 100,000 by 100,000 pixels from UCSF to solve medical problems.• Employing PyTorch-Lightning and Ray frameworks for large scale machine learning models hyperparameters tuning parallelly on GPU clusters.• Using deep learning methods and physical modeling of amplification process in Polymerase Chain Reaction (PCR) test to improve accuracy and efficiency of COVID-19 PCR test and understand contributing factors to the amplification efficiency.

• Developed and optimized a 2-step algorithm to verify summary statements generated from ChatGPT against original documents that increased verification rate of the existing pipeline by 5% while maintaining the precision in verified statements at 93%.• Created a high-quality dataset to benchmark future approaches to evaluate AI generated summaries.

• Implemented Vision Transformer model from HuggingFace to classify commercial building roof’s conditions using aerial images that performed better with existing models on classification metrics and covered more image types.• Built an algorithm that gives reasons for the model’s prediction using Auto-Grad attention heatmap and roof metadata.• Accelerated the model’s running time with ONNX runtime framework dynamic quantization technique.• Supported model’s deployment by wrapping the model with MLFlow and log it on DataBricks.

• Worked with physicians to extract, pre-preprocess and convert International Code of Diagnosis (ICD) from Contra Costa Health Center’s database using rxNorm API.• Constructed and analyzed temporal relationship between patient medical history (e.g. doctor visits, prescriptions, and diagnosis, etc.) and electrocardiogram (ECG) to identify occurrence of silent heart attacks.• Built a pipeline to run multiple deep learning models that are known to work well on ECG (e.g. LSTM, ResNet, Transformer) to predict historical silent heart attacks.

• Queried inventory data and sales transactions of 1000 + items over 10 years and automated weekly update process using SQL.• Explored and applied statistical and machine learning methods (e.g. time series, survival analysis, random forest) to predict future inventory levels for fast selling and slow selling products.• Analyzed products performance and warehouses sales; planned operational and pricing strategies with stakeholders based on the analysis reports.• Built and automated weekly sales and inventory reports for 200+ warehouses and monthly sales forecast using Python, R, SQL and VBA.

• Extracted and pre-processed California’s catchments attributes and pesticide usages from the United States Environmental Protection Agency and the California Department of Pesticide Regulation databases using arcGIS and R.• Explored methods to impute missing and censored data using spatial and temporal models.