Automating Regression Detection, Triage and Redeployment of Performance Monitoring VMs on the Azure Compute team- Deployed an automated system to periodically monitor virtual machine (VM) performance metrics and identify and report any incidents using the Kusto Query Language (KQL), and Microsoft Power Automate.- Employed database query optimization techniques for an order of magnitude improvement in speed.- Produced an algorithm that detects three types of regressions based on statistics.- Developed another algorithm that detects whenever a metric regresses and then improves using convolution.- Built a pipeline for automated redeployment of regressed or missing VMs using PowerShell, YAML, and KQL.- Detects faulty VMs with missing or regressed data 19 hours quicker and saves up to 2 developer hours per incident on critical VM identification, automated triage, and redeployment.

Deep learning-based computer vision for Pakistan real estate tax collection- Created a computer vision system to identify real estate features for accurate tax collection in Pakistan.- Employed transfer learning on state-of-the-art models (ResNet and ConvNeXT) using PyTorch andTensorFlow in Python and Jupyter Notebooks, increasing performance by more than 10%.- Constructed ensemble models that seamlessly combine multiple DNN architectures (ResNet-18, ResNet-34, ConvNeXT), which were then deployed on an app for the partner tax collectors to use in the field in Pakistan.- Vastly improved the existing deep learning systems with class balancing, sampling, data cleaning, and feature engineering with Pandas and Scikit-learn.

- Developed a Machine Learning-based noise detection solution using derived statistical features.- The project has the potential to save ~1400 device-hours per month, a 2.8% resource usage improvement.- Composed an in-depth, 13-page software design that also compared 16 different models.- Built an End-to-End ML system with Python and SQL to predict if performance regression detection parameters are properly set for a metric.- Trained multiple classification models with an internal AutoML framework using TensorFlow - including Random Forest, Decision Tree, and Neural Network models - and employed different training strategies.- Invented two original algorithms that validates and refines model output, eliminating 60% of noisy predictions.- Executed the full MLOps practice, from scoping to model deployment, in iterations.- Collaborated with another intern, who integrated the input and output of the ML system into the existing infra.

- Investigated the features crucial for cryptocurrency evaluation using data science techniques such as data collection, analysis, visualization, and feature engineering. Used Python, AWS, and basic SQL.- Led the crypto ledger analysis project, collaborating and communicating with members in other teams.- Contributed to approximately 50% of the company’s minimum viable product that will allow them to launch their business, as one of the first interns at this early-stage startup.

- Collaborated with two other students to research and engineer a 26-language compatible AI chatbot using the GPT-3 language model and DeepL translator APIs, with currently about 88,333 tokens of user input into the language model.- Researched a Japanese specific chatbot that would be more tailored for Japanese language learning. Showcases Presented at: - April 20, 2022 - Global Languages Student Research Showcase 2022- May 10, 2022 - SHASS Undergraduate Research Showcase - Spring 2022