Gainesville, FL —
Computer Science + Business Administration @ The University of Florida
currently
Hey! I'm Aryan. I'm a CS + Business student at the University of Florida.
Besides my professional and academic pursuits, making memories with friends and family is really important to me. I also like to play all kinds of sports, especially golf and pickleball. I love to eat amazing food (and now I'm learning to cook to satiate those cravings).
Always open to connect, chat, and explore opportunities. Feel free to reach out :)
clubs, research, and everything next
AlgoGators
Junior Quantitative Research
Florida Tech Pathways
Analyst
UF Data Science and Informatics
Sponsorship Coordinator
UF Data Studio
Undergraduate Research Assistant
Extern + Amazon
People/Data Analytics Extern
Mathnasium Learning Center
Mathematics Tutor
Florida State University Young Scholars Program
Research Assistant
a few things i've built
The Gap
Polymarket earnings trading strategy — 2.08/1.46 Sharpe, +8.67%/+8.29% across 2 exits.
Echo
Full-stack GenAI dashboard analyzing Google Reviews with a dual-model NLP pipeline.
House Pricing Multi Linear Regression
ML regression pipeline predicting house prices with SHAP explainability — R²=0.58.
Sharpe Strategy
Cross-asset momentum trading strategy — 0.9 Sharpe ratio, 17% portfolio growth.
Senti
CLI tool explaining stock movements using FinBERT sentiment analysis on real-time news.
Climbing the Language Tree
Cross-linguistic statistical analysis of pause patterns across 5 languages — R²=0.72.
other stuff
I grew up in [your city], where I first got into computers by [something interesting about how you got into tech]. Before UF, I [relevant backstory — school, projects, anything interesting].
Outside of tech I'm into [hobbies, interests, things you actually care about]. I find that the best ideas usually come from places that have nothing to do with computers.
At UF I've been [clubs, activities, things you're proud of that aren't pure résumé material].
Languages
Frameworks & Libraries
Tools & Platforms
Domains
Coursework
Identified Full-Time Warehouse Associates as the top attrition risk cohort out of 5 segments by building a Python (pandas, TextBlob) pipeline to clean and sentiment-score 200+ Glassdoor and YouTube reviews.
Built an Impact × Severity × Size scoring model to rank employee segments by attrition risk, surfacing burnout and favoritism in advancement as the core drivers.
Traced burnout to structural causes in work design using 5 Whys root cause analysis, then delivered findings in a pitch deck adopted for the intervention roadmap.
Scaled LLM inference across 10,000+ adversarial multi-agent simulations by integrating 20+ transformer models (7B–80B parameters) with 4-bit quantization and SLURM multi-GPU scheduling on HiPerGator HPC.
Achieved 85–90% F1 deceptive-agent detection across Logistic Regression, SVM, and MLP classifiers by building a scikit-learn NLP pipeline with TF-IDF vectorization and NLTK preprocessing over 289,000+ labeled utterances.
Built an interactive research dashboard with Flask and JavaScript for real-time ML suspicion scoring and agent visualization, enabling live behavioral monitoring across 20+ LLM configurations with sub-500ms state refresh.
Improved average student test scores by 18% over 4-month periods by adapting instruction style to individual learning needs across 250+ students from elementary through college levels.
Collaborated with instructors and center leadership in a team-teaching environment to relay student progress, identify knowledge gaps, and coordinate lesson adjustments, maintaining 90%+ family satisfaction across a 2-year tenure.
A volatility-adjusted momentum trading strategy achieving a 0.9 annual Sharpe ratio and 17% portfolio growth across 25+ liquid assets using inverse volatility weighting and 63-day ranking windows.
Backtested over 10 years of varied market conditions with realistic execution constraints — 5% stop-losses, monthly rebalancing, and 90/10 capital allocation — across equities, ETFs, commodities, and currencies.
An earnings-reaction strategy built on the gap between Polymarket's implied probability of a company beating earnings and that company's own trailing 8-quarter beat rate. A large disagreement between the two predicts how the stock moves the day after the report — evidence held up across 14,477 historical earnings events and 671 real Polymarket-matched contracts.
Backtested Oct 2025–Aug 2026 across 66 S&P 500 companies (94 trades, 49 long / 45 short) with asymmetric long/short thresholds, achieving a 2.08 Sharpe ratio and 8.67% return with a −2.00% max drawdown, validated against a 5,000-path Monte Carlo simulation.
A web app that lets you look up any business, pull its Google Reviews, and automatically analyze them with AI. Each review is scored for sentiment with a machine learning model, then a separate AI extracts complaints, praises, employee mentions, and suggestions. Everything shows up on a live dashboard with charts and cards.
Multiple users can share the same business data so reviews only get fetched and analyzed once — built with a multi-tenant shared data architecture using a PostgreSQL join table.
A machine learning model that predicts how much a house will sell for based on features like size, location, number of bedrooms/bathrooms, and condition. Cleaned and prepared a real dataset of ~4,000 U.S. home sales, engineered new features, trained a linear regression model, then used SHAP to explain which features matter most.
Found that Seattle location and above-ground living ratio are the top 2 price drivers, outranking raw square footage across all 132 features. Achieved R²=0.58 and MAE of ~$117K.
Actual vs Predicted
SHAP Features
Correlation Heatmap
Residuals vs Predicted
The model tracks actual sale prices well for homes under ~$750K, with dots clustering tightly around the ideal diagonal. Above $1M, predictions consistently fall short — the model underpredicts luxury homes.
A command-line tool that explains why stocks are going up or down. You type in stock tickers, and it pulls recent news articles, runs them through FinBERT (a transformer trained on financial text) to score sentiment, then compares that sentiment to the stock's actual price movement over the past week.
Uses exponential decay weighting (e^(-0.2t)) so today's news carries full weight and a 7-day-old article counts for about 25%. Outputs a plain-English interpretation — like "news-driven rally" or "technical rebound" — for each ticker.
Accepted into the Young Scholars Program at Florida State University. My team and I conducted a cross-linguistic statistical analysis of speech pause behavior under Dr. Michael Kaschak. We analyzed 1,000+ audio recordings across 5 languages (English, Spanish, French, Greek, Hindi).
We found that linguistically related languages share similar pause duration patterns. This was validated with regression models achieving R²=0.72 and p<0.01.
the team!
collecting data!
lunch break!
