Anaya
Gandhi
—

The Problem

Because proteins are inherently dynamic, mapping their conformational transitions is essential for modern drug discovery. However, traditional molecular dynamics simulations are computationally exhaustive, often requiring weeks to generate a single trajectory. This opens a critical question: can generative models bypass these physical simulation bottlenecks by learning the conformational landscape directly from historical trajectory data?

The Approach

This project implements a physics-informed VAE trained on MurD ligase molecular dynamics trajectories, consisting of 2,700 total frames evenly distributed at 900 frames per functional state (open, closed, and apo). The network utilizes an MLP encoder paired with a FoldingNet-style decoder, optimized via a total loss function defined as L_total = L_MSE + βL_KL + w_geoL_geo + w_ramaL_rama. This formulation directly accounts for reconstruction error, KL regularization, AMBER-style bond/angle penalties, and Ramachandran torsion constraints to preserve the physical realism of backbone dihedrals.

The optimal architecture was determined through a three-stage search evaluating 192 hyperparameter configurations. Test 15 established the baseline structural parameters (d_z=8, β=0.05, w_geo=5×10⁻⁵), while Config 82 identified the superior physics constraint set (bond+smoothness) with a realism score of 0.4965. Because this initial configuration exhibited mode collapse during structural interpolation, a final balanced VAE (d_z=64) was

Results

The Problem

For up to 80% of ALS patients, one of the earliest and most challenging symptoms to emerge is dysarthria the impairment of speech caused by the degeneration of motor neurons. Because this speech decline often appears long before other major physical symptoms, it presents a unique opportunity for early detection. Could automated speech analysis step in to catch these subtle vocal shifts before noticeable symptoms take hold? Transitioning from manual clinical evaluations to automated, data-driven speech tracking could completely change how early we catch the disease.

The Approach

Two parallel pipelines: Wav2Vec 2.0 fine-tuned on TORGO — a 2,000-sample database of dysarthric and healthy speech, evenly split by gender — for deep feature extraction and binary dysarthria classification. In parallel, a shallow neural net trained on 13 MFCCs served as a comparator classifier on the same dataset.

The second pipeline targets ALS severity prediction using the VOC-ALS dataset (153 Italian-speaking individuals: 102 ALS patients, 51 healthy controls). Five acoustic biomarkers — mean F₀, F₀ standard deviation, HNR, jitter, and shimmer — were extracted across 8 prescribed vocal tasks (40 features total) and fed into regression models: Linear, Ridge, Lasso, XGBoost, and Random Forest. Severity was quantified via an inverted ALSFRS-R clinical scale, enabling disease progression tracking.

Results

The Problem

Why do certain tracks peak immediately and fade into obscurity, while others build a slow, sustainable burn? This project investigates whether a song's full popularity trajectory. From its initial release to its peak and eventual decline, can be predicted using a combination of intrinsic audio features and strategic release timing.

The Approach

Using the Spotify Top 200 dataset (2017–2023) supplemented with Spotify Developer API metadata, I engineered three core lifecycle features: time to peak, duration at peak, and rate of decline. The feature set integrated temporal markers (release month and day of week) with core audio dimensions, including energy, danceability, valence, speechiness, acousticness, and loudness.

Unsupervised benchmarking across four clustering algorithms (KMeans, DBSCAN, Agglomerative, and Gaussian Mixture) evaluated via Silhouette, Davies-Bouldin, and Calinski-Harabasz metrics identified KMeans as the optimal model (Silhouette score of 0.214 vs. DBSCAN's −0.283). The elbow method established 3 distinct clusters as the ideal structural fit. Predictive modeling using Random Forest, XGBoost, and SVR demonstrated that incorporating release-month temporal features drastically outperformed models utilizing audio features alone when forecasting a track's time-to-peak.

Results

The Problem

Traditional financial analysis often overlooks the fact that stock correlations are heavily driven by shared investor behavior rather than fundamental market indicators alone. When the same institutional investors simultaneously hold large positions in Microsoft (MSFT) and another equity, broader market movements inevitably propagate through these underlying ownership networks. This structural interconnectedness remains completely invisible when examining price data in isolation, creating a critical blind spot for risk assessment.

The Approach

To uncover these hidden dynamics, this project constructed a bipartite graph utilizing quarterly 13F holdings filings spanning from 2016 to 2023. The resulting network comprised 1,606 total nodes—consisting of 77 distinct institutional investors and 1,529 individual stocks—with edge weights mathematically encoding the exact number of quarters a given investor maintained a position in a stock. To isolate the relationships between assets, this bipartite network was then projected onto the stock side, creating a unipartite co-investment network where an explicit link is established between two stocks if they share at least one mutual institutional investor

Three analytical threads were executed in parallel to evaluate this network structure. First, betweenness centrality was calculated to track which specific stocks acted as critical systemic bridges across the market. Second, Louvain community detection was applied across 29 unique quarterly snapshots to monitor how clusters of co-invested stocks evolved over time. Finally, a minimum spanning tree was constructed based on historical return correlations, using the distance formula d = 1 − |correlation|, to reveal the market's true structural backbone before, during, and immediately following the COVID-19 pandemic.

Results

The Problem

Market dynamics really come down to what happens when different trading strategies collide. When you have trend-followers, mean-reversion traders, and noise traders all competing in the exact same space, it begs the question: who actually comes out on top, and what does that competition do to the market itself? Instead of using rigid, top-down economic equations, agent-based modeling lets these complex behaviors emerge naturally from the bottom up. This makes it much easier to actually see and map out exactly how things like sudden volatility cascades and wealth concentration happen.

The Approach

To test this out, I set up a simulation of a Bitcoin/GBP market running over 1,750 days. I started with a baseline population of 100 agents, giving every single one of them a starting balance of 5,000 GBP and 0.5 BTC to keep the playing field level. For the initial population mix, I went with 70% Chartists—which included a mix of Momentum-Dominant, RSI-Dominant, Balanced, and Alternating subtypes—and 30% Random Traders. After establishing that baseline, I threw a brand-new Contrarian Trader class into the mix, replacing 10% of those Random Traders to see if a mean-reversion strategy could actually step in and stabilize the market.

As for how the market handles the math, pricing is driven by Petrov's Price Response Function, meaning net demand is what directly forces price changes. On the supply side, Gibrat's Law dictates Bitcoin's growth, kicking out a +10% supply increase every 90 days that gets distributed to agents based on how much crypto they already hold. To keep things realistic and mimic real-world portfolio rebalancing, I built in a market correction mechanism that triggers every 200 days. Finally, after tuning the parameters, I found that an optimal trend-length of n = 10 worked best across all the different chartist subtypes.

Results