Co-Researchers: Moin Mattar, Enrico Rodriguez

PyTorch · Diffusion Transformers · VAE · WebSocket

A diffusion-based world model to simulate 2D billiards (pool) physics with no physics engine at inference. An action-conditioned DiT predicts each frame in VAE latent space, using diffusion forcing for long-horizon stability. Trained on a 200,000-episode dataset and playable in a real-time WebSocket neural simulator.

LISP Lab, Dartmouth · Co-Researchers: Peter Chin, Yu-Wing Tai, Abhinav Reddy

PyTorch · Mamba / SSMs · CUDA · Python

An extension to Mamba's state-space model that adds embedding search and injection to improve long-sequence recall, matching Transformer baselines in linear time. Embeds the query and input tokens, runs a similarity search for relevant context, and injects those hidden states into the current state rather than processing the full sequence memory.

LISP Lab, Dartmouth · Co-Researchers: Peter Chin (advisor), Minh Bui (PI), Paul Cherian, Mary Wood

PyTorch · DeepSpeech2 · Federated Learning · CTC Loss

An adversarial attack that reconstructs private voice samples with 99% fidelity from the gradients of speech-recognition models. Reverse engineers CTC loss in federated learning environments, recovering original training audio & validated on DeepSpeech2.

code ↗ · poster ↗

First Author · Co-Researchers: Sanam Mittal

Python · scikit-learn · NLP · PyTorch

A non-invasive method for early dementia diagnosis that pairs NLP with machine learning. Instead of brain scans or blood tests, it derives diagnostic signals from cognitive and language assessments, enabling accessible screening for elderly populations. Built during research at Optum, published at ICDSCA 2022.

First Author · Journal of Emerging Investigators

Self-driving cars go blind in bad weather. A two-stage deep learning system fixes that: one module masks and removes rain, snow, and fog, then a detector labels the cleaned frames — 89.72% average accuracy, beating industry baselines on both accuracy and detection time.

Dietlein Lab, Harvard Medical School · Co-Researchers: Prof. Felix Dietlein

Reading single-cell RNA-seq to find what drives tumors. Clustering and differential-expression analysis surface the gene shifts behind malignant transition, and a random-forest pipeline predicts ligand–receptor interactions at those sites at 89% accuracy to guide therapeutic targeting.

code ↗ · paper ↗

Dartmouth Economics · Co-Researchers: John Welborn, Victor Van Erp

Python · scikit-learn · Financial Engineering · NumPy

A pipeline to predict next-day equity-ETF downturns from implied-volatility skew between deep out-of-the-money and at-the-money puts as the signal. Trained on 1.77M QQQ option records across the 2020 crash and 2022 correction, the ensemble reaches 0.91 AUC at sub-50ms latency.

code ↗ · paper ↗

independent · first author · Co-Researchers: Amit Baranwal (DataCore), Sugandha Sahay (CMU)

How automated is a company's testing, really? This system authenticates against a company's cloud, pulls test-case records through the Jira and Zephyr Scale REST APIs, and turns them into clean automation-coverage and status metrics for management.

Dartmouth Economics · Co-Researchers: John Welborn, Victor Van Erp

A pipeline to predict next-day equity-ETF downturns from implied-volatility skew between deep out-of-the-money and at-the-money puts as the signal. Trained on 1.77M QQQ option records across the 2020 crash and 2022 correction, the ensemble reaches 0.91 AUC at sub-50ms latency.

code ↗ · paper ↗

Python · NumPy · SciPy · Financial Engineering

Real-time calibration algorithm for the Heston stochastic-volatility model on live SPY/SPX option surfaces. Priced via the characteristic function with a Carr–Madan FFT, it uses a four-stage optimizer that reaches an R² of 0.94 and recalibrates in under 20ms.

Python · Streamlit · scikit-learn · NumPy

An educational tool I built for Dartmouth’s Mathematical Finance I as TA. Quantra pulls live market data and prices contracts with Black–Scholes, a Cox–Ross–Rubinstein tree, and a 10,000-path Monte Carlo. Also computes analytical and numerical Greeks, delta-neutral hedges, and full implied-vol surfaces.

code ↗ · demo ↗

AI agents that eliminate infrastructure waste

React · TypeScript · Python · LangGraph · MCP

A startup I co-founded with other AWS interns in Summer 2025. Martian applies agentic AI to enterprise infrastructure cost optimization, with agents map storage dependencies and execute runbooks to eliminate wasted spend. A LangGraph multi-agent MCP server drives the analysis with integrations to Azure, AWS, & GCP cloud environments, fronted by a TypeScript interface for fintech teams.

Live test-automation coverage dashboard

Python · FastAPI · React

Novel methodology to measure test-case automation across corporate software infrastructure. Authenticates against a company’s cloud servers, pulls test-case records, and parses them into metrics on automation coverage and status for management. Published & presented at GTBMC 2022.

Review rating and category classification

Python · scikit-learn · Jupyter

A machine learning pipeline that classifies Amazon reviews by rating and clusters them by category with scikit-learn. TF-IDF and SGD classifiers tuned per rating cutoff, beating Kaggle baselines at every threshold, while clustering runs on 40 binary category-keyword features. Built for Dartmouth CS74.

Real-time multiplayer terminal mining game

C++ · UDP / sockets · Valgrind

A real-time multiplayer terminal mining game where 26 players navigate a maze for gold while an authoritative server syncs clients over UDP, with ray-cast line-of-sight visibility. Originally built for CS50 Final in C, the C++ rewrite swaps manual memory management for RAII and an unordered_map grid, rendering the full map in under 2ms per frame.

A tiny search engine: crawl, index, rank

C · HTML · Valgrind

A search engine in C built as three programs that communicate through the file system: a crawler that walks from a seed URL to a set depth and caches pages, an indexer that builds a word-to-page index, and a querier that evaluates AND/OR queries into ranked results. Built for Dartmouth CS50.