Silver Tech Studios
Software Development Engineer / AI Engineer
At Silver Tech Studios, I develop AI-powered software that streamlines HSA/FSA workflows for both businesses and consumers.
B2B Auto-Adjudication Platform - Silver AI Claim Agent
I build intelligent document processing systems that automate HSA/FSA claims adjudication for third-party administrators (TPAs). My work includes developing OCR and information extraction pipelines, extracting structured information from medical receipts and invoices, and building machine learning models that determine HSA/FSA reimbursement eligibility. These systems reduce manual review while improving the speed and accuracy of claims processing.
Consumer Platform - WithSilver.app
I also contribute to Silver's consumer platform, which helps users maximize the value of their HSA/FSA benefits. I develop data pipelines that helps process data gathered from major retailers such as Target and Walmart to identify products eligible for HSA/FSA reimbursement. By combining large language models with custom classification models, we automatically surface reimbursable purchases that users might otherwise overlook, helping users save both time and money.
Weather App
Project Description
The Weather App is an Android application designed to deliver accurate and current weather information to users. It utilizes tomorrow.api to fetch real-time weather data for various locations, ensuring a comprehensive user experience.
Key Features
City Search: Users can swiftly search for cities with a search bar featuring an auto-complete option for quick and easy location selection.
Weather Display: Once a city is chosen, the app retrieves and presents relevant weather data in an intuitive format that is easy to read and understand.
Visualization: The app employs Highcharts to create visually appealing charts that illustrate weather trends, enhancing user comprehension of the data.
Social Integration: Users can share current weather conditions directly to X (formerly Twitter) from within the app, promoting social engagement and information sharing.
Technology Stack
The Weather App is developed using the MEAN stack, which includes:
MongoDB: Used for data storage, ensuring that user and weather data is managed efficiently.
Express.js: Serves as the web application framework, facilitating the handling of server-side logic.
Angular: Powers the front-end user interface, providing a responsive and interactive experience for users.
Node.js: Acts as the server-side environment, enabling scalable network applications capable of handling multiple requests simultaneously.
This robust combination of technologies allows the Weather App to deliver real-time weather updates effectively and efficiently to users.
Demo of the core features of the app
SWEET - Weakly Supervised Person Name Extraction for Fighting Human Trafficking
SWEET (Supervise Weakly for Entity Extraction to Fight Trafficking) is a weak supervision framework designed to extract person names from noisy and unstructured escort advertisement data.
A key challenge in this domain is the lack of labeled training data, as well as the highly variable and adversarial nature of the text. We also study large language models in the context of the human trafficking (HT) domain, with a particular focus on their behavior under adversarial prompting. As an aside, we examine how models may produce harmful, toxic, or explicit content when subjected to jailbreak-style inputs used to elicit synthetic HT-domain data. This analysis helps motivate the need for robust filtering, safety considerations, and weak supervision strategies when working with real-world scraped and model-generated data.
To address the low-resource setting, SWEET combines multiple weak supervision sources, including rule-based labeling functions (such as pattern-based rules and “antirules” for negation) and labels generated by large language models trained on domain-specific and synthetic datasets.
We also introduce HTGen, a synthetic dataset generated using a large language model to simulate structured escort advertisement data, which improves robustness and coverage across diverse linguistic patterns.
The weak supervision framework aggregates these noisy signals into a unified probabilistic training signal, enabling improved entity extraction performance in challenging, high-noise environments.
This project was conducted under the supervision of Professor. Reihaneh Rabbany while I was working in the Complex Data Lab and on the Mila Infrared Project
Image Classification from Scratch and Scaling Using Multicore Matrix Multiplication
During my experience as a research assistant at the Prometheus Lab at McGill university, I engaged in a project focused on image classification from scratch. The goal was to develop a Java-based vision system that could effectively classify the type of room a robot was navigating.
The project involved constructing a custom image classification program, which was built entirely from the ground up. To optimize the handling of images captured by the robot, I implemented a batching system. This approach allowed for efficient processing of multiple images simultaneously, reducing overall running time.
I also integrated serialization in the codebase, facilitating the saving and loading of both the model and training data. This enhancement significantly improved the workflow, enabling easier management of the various datasets used during training and testing phases.
To scale the application for larger datasets, I utilized multicore matrix multiplication packages. This approach leveraged parallel processing capabilities, which enhanced computational efficiency and supported the handling of increased volumes of training and testing data.
The culmination of these efforts resulted in a robust image classification system adept at rapidly and accurately identifying the type of room based on visual inputs from the robot, reflecting the project's objectives and the practical application of advanced programming techniques in machine learning.
Brain Tumor Segmentation with Attention-Based U-Net
This project focused on enhancing brain tumor segmentation through an improved U-Net architecture. The primary modifications involved the integration of Squeeze-and-Excitation Blocks and Convolutional Block Attention Modules (CBAM) into the decoder sections of the original U-Net model.
The Squeeze-and-Excitation Block enhances the representational power of the network by adaptively recalibrating channel-wise feature responses. This approach allows the model to focus on more relevant features while suppressing less informative ones, thus facilitating better segmentation outcomes.
Similarly, the CBAM attention module introduces a dual attention mechanism, which sequentially infers attention maps along the channel and spatial dimensions. This helps the model to better emphasize critical regions within the input images that contribute significantly to effective segmentation.
The combination of these attention mechanisms within the U-Net decoder has demonstrated noteworthy improvements in performance metrics on the brain tumor segmentation task, showcasing the efficacy of incorporating advanced attention techniques in neural network designs. The research findings from this project were later published as a full-length paper, contributing to the ongoing exploration of deep learning applications in medical imaging.
Rocket Engine DAQ
During my internship at USC Liquid Propulsion Laboratory, I had the opportunity to contribute to the Data Acquisition (DAQ) team. My primary focus was on implementing Python and Labjack code that facilitated the retrieval and visualization of engine data (temperature, force measure, etc). This data was collected from Labjack devices integrated within the DAQ monitoring system.
My responsibilities included developing scripts that would efficiently fetch real-time engine data, ensuring accuracy and reliability. I utilized libraries such as Matplotlib and Pandas for data visualization, enabling our team to analyze performance metrics effectively. This experience not only enhanced my programming skills but also deepened my understanding of the complexities involved in monitoring rocket engine performance.
Collaboration with team members was crucial as we worked together to troubleshoot any issues that arose within the data collection process (issues with the timing library in python that lead to inaccurate launch sequence). Overall, this internship provided me with a solid foundation in data acquisition systems and their application in aerospace engineering.
More videos about the hot fire HERE
