Hello, my name is Mohsin Shah. I'm a computer science and math double major at the University of Massachusetts Amherst.
I'm particularly interested in AI/ML, especially in natural language processing and multimodal models. My focus has been on developing systems that generate prompts or captions from AI-generated images. At Microsoft, I implemented customizable benchmarking metrics, error analysis modules, and LIME explainers for LLMs. Currently, I'm learning more about optimized models for tabular data analysis.
Additionally, I like running, Brazilian jiu-jitsu, and photo/video editing.
Experience
Fidelity Investments
Data Engineering Intern
• Created chatbot for single sign-on service with Amazon Lex, tracked via Jira; aiding efficiency and projected to reduce inquiries by 40%.
• Designed Splunk dashboards for Password Resets & User Registration, identifying friction and abandonment points for millions of users.
• Applied Snowflake SQL tables & AWS S3 to migrate client-facing cybersecurity data, raising data security for 5000+ annual presentations.
Microsoft
Data Science Intern
• Extended Azure ML’s Responsible AI Toolbox & Interpret Text for LLMs like GPT-4 & Llama, aiding 200,000+ users in model evaluation.
• Implemented LIME explainers, customizable benchmarking metrics, and error analysis modules in the comprehensive UI dashboard.
• Developed 5 tutorial notebooks showcasing model analysis with HuggingFace (GPT-Neo, RoBERTa) and OpenAI API (GPT-4, 3.5, 3).
University of Massachusetts Amherst
ML & NLP Research Intern
• Analyzed multimodal transformer models: BLIP, GIT, CLIP, and custom vision language model (VLM) with BERT (LLM) encodings, EfficientNet (CNN), and LSTMs with PyTorch (CUDA) to generate prompts of AI generated images, achieving a BLEU score of 68%.
• Created training and validation datasets for R&D using Python & Selenium, web scraping 1000+ AI generated images and prompts.
Biologically Inspired Neural and Dynamical Systems Lab
AI & RNN Research Intern
• Built simulations in Julia to study the applications and dynamics of oscillatory neural networks; made computation 10x faster.
• Designed algorithms to solve the ongoing challenge of recurrent neural network oversaturation; potentially applicable in robotics.
• Enhanced data visualization with 1000+ raster plots and video heatmaps, integrating clustering algorithms for data segmentation.
Projects
Sign Language AI Translator
Sign Decoder is an AI system that translates sign language to text and speech in real time. The goal of our app is to be accessible and free, bridging the gap between signers and non-signers. Our app is going to have an intuitive and friction-less design, simply point the camera at the person signing to start translating! Our prototype was awarded "Best use of an AI model" by travelers.com at the Hack(h)er 413 Hackathon (2023).
eBay: ML & NER Competition
We Developed a 94% accurate name entity recognition (NER) model using 10 million raw eBay listings in German; effectively classifying each word. To do this we analyzed and preprocessed the raw, non-english dataset with Pandas; streamlining feature extraction and performance. Ultimately we achieved our goal of enhancing the data quality and searchability of the eBay listings.
ShareSpace: Find Roommates
I collaborated with a team of 10 to develop a full-stack web app that matches roommates based on their preferences, allowing matched users to chat and customize their profiles. The goal of this project was to create a centralized roommate finder for UMass students. I learned a lot while working on this project, especially the specifics of building and interacting with databases.
Automated Video Content Creation App
We developed an app that allows users to create social media videos by selecting book segments and narrator voices. This project automated the entire video creation process, including syncing background clips and on-screen text. As a result, we currently have over 5,000 followers.
Customer Churn Prediction
This project focuses on predicting customer churn for a telecommunications company by identifying key drivers and delivering actionable insights. XGBoost was selected for its performance and scalability, with SMOTE addressing the dataset's imbalance. Using XGBoost's feature importance and SHAP explainers, we derived insights that can inform strategies to enhance customer retention and profitability.
AI Flappy Bird
I made this Flappy Bird AI by first applying object-oriented programming to make the general mechanics of Flappy Bird with Python and PyGame. While developing the game, special attention was given to simplifying the simulation of physics and collisions. Then I implemented the NEAT (NeuroEvolution of Augmenting Topologies) genetic algorithm to create intelligent, evolving birds that can play the game autonomously. As a result, the AI birds were capable of playing the game indefinitely by the 11th generation.