*Click here to see project details.
My team and I investigated the impact of subsidy policies on EV adoption in China using city–month data from 15 cities between 2016 and 2019. We applied panel regression and instrumental variables—specifically the local-to-national subsidy ratio—to address endogeneity. Results show that a 10,000 Yuan increase in subsidies boosts EV sales by 6–11%, controlling for GDP, income, and charging infrastructure. We chose this period for its stable, consumer-focused policies. While subsidies significantly increased adoption, we also found that local industrial dynamics and consumers’ expectations of future policies influenced results. Due to limited data granularity and transparency, we recommend future research explore interactions among multiple policy tools and variations in consumer responses.
*Click here to see the published paper.
My team and I researched how hog futures could reduce price volatility in China’s pig market, which has long faced cyclical fluctuations. Using the herding effect and cobweb model, we showed how irrational decisions and supply lags drive instability. We argued that hog futures provide essential functions—price discovery, hedging, vertical integration, and support for large-scale production—that can stabilize the market. Through theoretical analysis and a case study of Muyuan Food Co., we found that major producers began using futures to manage risk and inform production. However, high capital barriers and technical complexity limited small-scale farmer participation. Although China’s hog futures market remained nascent, we concluded it holds strong potential to enhance market stability, reduce volatility, and improve long-term risk management across the supply chain.
*Click here to see project details.
My friends and I built EcoBear in a 2-day hackathon, a web-widget to bridge the gap between consumers’ sustainable intentions and their actual shopping behavior. Most shoppers wanted to buy green but lacked clear, trustworthy information and were overwhelmed by greenwashing and convenience. Our tool integrated directly into shopping websites, automatically scanning products and displaying instant sustainability ratings using color-coded labels—green for sustainable, red for unsustainable. Users could click to learn more about certifications and evaluation criteria.
To drive engagement, we gamified the experience with rewards and a friendly polar bear mascot. Eco-conscious users appreciated the badges and rewards, while casual shoppers responded well to clear alternatives and incentives. From our testing, we learned to make sustainability cues more prominent and reduce guilt-driven nudging.
*Click here to see project details.
As a participant in the WiDS Datathon 2025, I was tasked to analyze fMRI brain imaging data to uncover how biological sex influenced the presentation of ADHD. Me and my teammates analyzed neuroimaging (fMRI) data from over 1,000 pediatric subjects to predict biological sex and ADHD diagnosis. We engineered advanced statistical features and applied principal component analysis (PCA) for dimensionality reduction. Using Python and R, we built and evaluated predictive models including convolutional neural networks (CNN), random forests, and logistic regression and contributed to advancing research into sex-specific neurobiological patterns associated with ADHD.
*Click here to see the archive.
I led a team of four to create a full hackathon platform from scratch — including the main site, past event archive, merch designs, and all visual assets.
*Click here to see project details.
I led a four-person team extending the CommonLit Readability Prize winning solution into a system that predicts text difficulty out-of-sample — scored only on corpora absent from training — after showing the unmodified pipeline's in-corpus Spearman of 0.891 collapsed to 0.647 cross-corpus. We built the evaluation harness before any model: rank-based metrics with entire corpora held out, every score bracketed between noise floors, formula baselines, and the contest's human-level ceiling, every comparison run at three seeds with paired-bootstrap confidence intervals, and the ensemble weight pre-registered before evaluation. The experiment platform is config-as-data — every variant is a config value, never a code edit — with isolated artifact paths, fail-fast guards, and a 34-test pytest suite. The measurements repeatedly overturned the inherited narrative: the headline pseudo-labeling mechanism measured null out-of-sample and was killed, label diversity proved the binding constraint (one added human-labeled corpus was worth 4–11× any architecture change), and the same formula signal hurt as a training feature yet won as a zero-training rank-aggregation ensemble. The highest-impact signal came from 20k Simple/standard-English Wikipedia article pairs fed through a group-masked pairwise-ranking head in a causally clean A/B, lifting held-out Spearman from 0.446 to 0.594 and collapsing seed-to-seed variance 6× — and decomposing that gain falsified our own hypothesis, tracing all of it to sharper cross-document ranking. Along the way we root-caused a DeBERTa-v3 first-step NaN blow-up via controlled backbone bisection and audited a silent text-windowing corruption, quantifying its effect on every conclusion rather than assuming. The final system beat classic readability-formula baselines on all three held-out evaluations (Spearman 0.89 vs 0.49 in-corpus; 0.59 vs 0.40 and 0.845 vs 0.79 cross-corpus).
*Click here to see the published paper.
As an undergraduate researcher at MIT CSAIL, I contributed to BigFANN (PVLDB 2026), a label-aware filtered vector search framework built on two ideas: a hybrid graph index whose heterogeneous edges — per-label exclusive edges guaranteeing connectivity inside every filter's subgraph (defeating the label interference that collapses mixed-label graphs' usable degree), plus shared edges reused across labels — tune to any memory budget between fully isolated per-label indices and a single unified graph; and join-free explicit filtering that probes only the smaller label's IVF clusters for double-filter queries instead of gathering and sorting candidates from both. The framework delivers up to 786× higher throughput than UNG (YFCC10M double-filter, 90% recall) and up to 4× over ParlayIVF2 at equal or better recall, across six datasets up to 100 million vectors. I ran the framework's experimental campaign on an NSF ACCESS allocation at the Texas A&M HPRC center: sweeping edge count and edge composition over the exclusive/shared grid — the sensitivity study behind the paper's parameter-selection rule and its memory-bound flagship result, where the tuned index fits 100M uint8 vectors in 64 GB at over 95% recall while index inflation forces ParlayIVF2 down to a sub-80% recall plateau — and characterizing filter behavior end-to-end: query specificity spanning four orders of magnitude to locate the IVF-to-brute-force crossover, Zipf/Poisson label distributions, and real workload splits across 200,386 distinct labels. I also profiled the query-time anatomy behind the performance analysis: per-stage breakdowns isolating the join pipeline's sort as the dominant overhead — ParlayIVF2 gathers and sorts a million candidates to reach 99% recall on SIFT100M double-filter — distance-computation censuses showing join-free search cutting 3,574 to 2,356 computations per query (YFCC10M double-filter, 99% recall), and Pareto-tuning of every baseline to its best possible showing. The C++ backend's memory-bound hot loops were optimized with software prefetching, insertion-based top-k maintenance, SIMD filter checks, vectorized uint8 distance kernels, and memory-bounded bitset filters.
*Click here to see project details.
I designed and built an end-to-end research and backtesting framework on QuantConnect/LEAN to test whether a regime-aware allocator over two classic signals — 12-1 momentum and short-term reversal — beats simple baselines out-of-sample on point-in-time, survivorship-bias-free US equities, prioritizing reproducibility, matched comparison, and honest significance over headline returns. The codebase is structured as modular LEAN algorithms computing both factors from one shared 252-bar rolling window, with every allocation rule behind a single extensibility hook and every research lever exposed as a named constant — so pure momentum and pure reversal fall out of the combined, dollar-neutral book as perfectly matched A/B baselines. When the backtester's 10,000-trade cap limited runs to about four years, I recognized it as an execution limit rather than a data limit and built an offline sleeve-reconstruction engine that rebuilds both factors' daily return series from history queries alone, unlocking the full 26-year (2000–2026) sample — streamed month-by-month to bound memory and gated behind a ground-truth test that had to reproduce the live backtests to the decimal before extrapolating. Every strategy is characterized net of transaction costs, with Probabilistic Sharpe for significance and a skew/kurtosis-corrected Deflated Sharpe applied across the ~26-configuration experiment ledger, all hyperparameters pre-registered and git-committed before running. The findings trace the edge to estimable quantities: volatility-scaling lifted the trend leg's Sharpe from 0.27 to 0.40, an anti-correlated reversal leg halved drawdown, naive return-timing whipsawed and failed, and the overall edge deflates to suggestive-but-not-conclusive over 26 years — reported as such, with the deliberate call to stop before over-fitting rather than climb model rungs the analysis showed would only degrade significance.