Mike Terekhov
MS CS @ University of Southern California
Boston University - B.S. Mechanical Engineering with a Concentration in Machine Learning
University of Southern California - M.S. Computer Science
Current Research Projects
HIV Research AI Regimen Project
I am developing an AI-powered clinical decision tool using Retrieval-Augmented Generation (RAG) to optimize Antiretroviral Therapy recommendations for HIV patients globally. My project features a round-table style AI framework where multiple AI agents simulate perspectives of virologists, clinicians, and pharmacists to collaboratively refine treatments. I am implementing agentic workflows with Agents, Planning, and Reasoning Chains to analyze drug-mutation interactions and generate personalized treatment recommendations.
ACL Injury Biomechanics AI Project
Using multimodal biomechanical sensor data collected across multiple clinical time points, you trained an LSTM-based classifier to distinguish between three groups: Control, Non-Injured, and Injured athletes in the context of ACL injury.
The goal of this project was to specifically see if we can identify between movement patterns of fully healthy legs versus the non-injured leg of athletes who have experienced an ACL injury. Using this knowledge, the model aims to detect subtle biomechanical differences that may indicate compensatory movement strategies or lingering deficits in the non-injured leg that could predispose athletes to future injury. With this data, we can help injured patients better understand their movement patterns and potentially guide rehabilitation strategies to reduce the risk of re-injury.
The model processes full temporal sequences of normalized features, learning patterns across time to classify each subject. Experimenting with varying numbers of training epochs and roughly 54,000 parameters, the model showed strongest performance on the Control class, while Non-Injured and Injured subjects proved harder to separate, a clinically meaningful challenge given how similar their movement patterns can be post-injury. Results were further broken down by clinical time point and sex to assess where the model generalizes well and where it struggles.
Previous Research Projects
AI vs Human Text Classifier
In this project, my team and I set out to explore whether machine learning models could effectively distinguish between human-written and AI-generated high school essays. We gathered a diverse dataset that included original human essays, AI-generated essays using models like GPT-4, GPT-2, Mistral, and Gemma, as well as both AI- and human-paraphrased texts. We applied classical text analysis methods such as Bag of Words, TF-IDF, and SVMs, alongside a neural network using GloVe embeddings. While traditional methods performed well on standard AI text, we found that they struggled with identifying AI-paraphrased versions of human essays. However, our GloVe-based model showed strong performance even in these more nuanced cases. Through this work, we demonstrated both the potential and the challenges in detecting AI-generated content, particularly as language models become more adept at mimicking human writing.
View Paper
LLM Text to SQL Model
For this project, we developed a text-to-SQL system by fine-tuning the DeepSeek-Coder 1.3B model using Low-Rank Adaptation (LoRA) on a curated dataset of 1000+ natural language and SQL query pairs from an NBA database. We integrated a Retrieval-Augmented Generation (RAG) module to dynamically improve prompts based on schema similarity, which helped optimize model performance. Through fine-tuning and data augmentation strategies, we increased SQL validity by 22% and result correctness by 23% over the baseline. I also built a custom evaluation framework to track SQL validity, result accuracy, and query matching across experiments.
View Paper
Vision Transformer Research
I am exploring methods to detect voids in Composite Oriented Strand Boards using machine learning, aiming for up to 95% accuracy. I formulated and implemented a Vision Transformer architecture with an encoder and 8 attention heads to explore advanced image representation techniques. My approach includes a preprocessing pipeline that segments 128×128 images into 16×16 patches for the Vision Transformer, enabling efficient processing of micro-CT scan data for void detection.