Daily long/short strategy on ~620 USDT-quoted Binance perpetual futures (2020 – 2026, 381k symbol-days). Pulled hourly klines and funding rates from a ClickHouse warehouse, built seven literature-backed factors as cross-sectional rank z-scores, validated them with Fama–MacBeth regressions, and traded a top-N / bottom-N portfolio rebalanced daily — benchmarking a transparent factor blend against a stack of ML models.

• Engineered the data pipeline against a remote ClickHouse binance database, pushing OHLCV / funding aggregation server-side (argMin / argMax / HAVING hours = 24) so 11M hourly rows never crossed the wire; cached as parquet and applied a $5M 30-day average dollar-volume filter to control for survivorship and liquidity.
• Constructed 7 factors — momentum (7d, 30d), 1-day reversal, funding rate, taker-buy imbalance, Amihud illiquidity, 30-day realized volatility — all standardized as daily cross-sectional rank z-scores under a strict no-lookahead invariant (factors at close of t, return t→t+1).
• Ran daily Fama–MacBeth regressions with Newey–West (lag 5) HAC standard errors; three factors cleared |t| > 2.5 — Amihud illiquidity (t = +4.36), funding rate (t = −2.96, sign consistent with basis-trade theory), and taker-buy imbalance (t = +2.83, the unique-to-Binance signal).
• Backtested a combined factor blend (top-10 / bot-10, equal-weight, 2× gross book) achieving Sharpe 1.89 gross / 1.28 net of 10 bps round-trip, +135% annualized gross return, max drawdown −60% across 2,048 trading days.
• Benchmarked an ML stack — OLS, RidgeCV, RandomForest, XGBoost, LightGBM, a small PyTorch DNN (9→64→32→1, BatchNorm + dropout), and an equal-weighted ensemble of the four nonlinear models — using expanding-window monthly refits, target winsorization, and a chronological 10% validation slice for early stopping; reported OOS R² in the Gu–Kelly–Xiu (2020) sense. The nonlinear ensemble reached Sharpe 2.14 gross / 1.23 net, beating linear baselines by capturing factor interactions (e.g., funding works harder on illiquid names).
• Conclusion: the simple combined-factor blend is the winner net of costs — transparent attribution and Sharpe 1.28 net — while the ensemble's higher gross edge is partly given back through turnover. Three factors carry most of the signal; ML's contribution is interaction terms, not net new alpha.

Differential expression analyses are susceptible to false positives. permuteDE uses permutation testing to identify which comparisons have a higher number of significant differentially expressed features than would be expected by chance.

• Developed the permuteDE R package, overseeing maintenance, debugging, and feature development.
• Collected and integrated eight large-scale public snRNA-seq datasets from Synapse, performing rigorous data cleaning, metadata curation, and structuring data into Seurat objects and BPCells matrices.
• Applied the permuteDE pipeline across eight datasets to demonstrate its real-world utility.

Note: Currently we are finalizing the package and analyses; permuteDE will be released upon manuscript publication.

Figure Source: Emani et al., Science 2024  ·  Pampari et al., bioRxiv preprint 2025  ·  Ameen et al., Cell 2022

I joined the PsychENCODE project at Prof. Anshul Kundaje's lab. I led this project analyzing scRNA-seq and scATAC-seq multiome data from three brain regions to investigate cell type-specific chromatin accessibility and gene expression patterns underlying schizophrenia and bipolar disorder to better understand their pathogeneses.

• Preprocessed high-dimensional multiome data through quality control, dimensionality reduction, clustering, doublet detection, marker identification, and cell type annotation.
• Performed differential gene expression & peak accessibility analyses (pseudobulk: Wald test; single-cell: Wilcoxon + MAST) across multiple cell types.
• Trained cell type-specific ChromBPNet models to identify causal genomic variants and their downstream effects on chromatin accessibility and gene regulation.

Figure Source: National Cancer Institute 2019  ·  Lowder et al., Cancer Letters 2022  ·  Langfelder & Horvath, BMC Bioinformatics 2008

Black men face higher prostate cancer incidence and mortality than White men. This project leverages LC-MS metabolomics on a cohort of 34 men (17 Black, 17 White) from the IRONMAN Registry to characterize metabolites associated with such disparities. I presented this project as my Master's thesis.

• Applied t-test, PCA, PLS-DA, random forest, logistic & linear regression, chemical similarity enrichment analysis, pathway analysis, and WGCNA to LC-MS metabolomics data, characterizing metabolites associated with Black-White prostate cancer disparities.
• Interpreted complex statistical findings and communicated them through clear scientific writing and effective data visualizations; defended before Stanford Epidemiology faculty and students.

This project visualized the interplay between life expectancy (measured by central death rates) and economic development, healthcare access, energy production, and consumption patterns.

• Developed interactive data visualization dashboards (global heatmaps, cross-country correlation analyses, and time-evolving bubble charts) using JavaScript D3.

Figure Source: Argelaguet et al., 2018  ·  Growing Life

In this project, I practiced NIH-style grant proposal writing by developing a theoretical proposal that integrates single-cell multiomics data to investigate the molecular mechanisms using factor analysis underlying Alzheimer's Disease.

• Drafted the Specific Aims section of a conceptual NIH grant proposal and presented the multi-omics factor analysis plan to ~30 professors and students.

This project evaluated associations between multiple exposures (age [categorical and continuous], sex, race, and smoking) and assessed alcohol as an effect modifier of the smoking–SAH relationship.

• Cleaned and managed case-control datasets in SAS OnDemand: resolved duplicate records, standardized missing values, and created derived variables.
• Conducted descriptive analyses (chi-square tests and t-tests) and applied univariate and multivariable logistic regression to estimate crude and adjusted odds ratios for smoking.
• Performed stratified analysis using Mantel–Haenszel testing to adjust for alcohol use as a confounder.
• Generated summary tables and data visualizations for reporting.

Nominated as Honorable Mention for Best UI for Hack Challenge Spring 2022.

This project developed an iOS app to ease the booking of a study space on campus in one of Cornell's many open libraries and study spaces.

• Programmatically developed a Cornell library study room booking system iOS app by integrating UIKit, AutoLayout, Navigation, UITable & UICollectionView, MVC, Delegation, and Animation.
• Implemented GET all libraries and available rooms, POST new reservation(s), UPDATE reservation history & DELETE reservation(s) that interact with backend API using Alamofire.
• Implemented UI with designers and collaborated with backend teammates for backend requests.