Projects

building a full-stack stock-research platform that combines market data, machine-learning signals, and portfolio analytics for S&P 500 companies. developed a time-series training pipeline with engineered technical, fundamental, and macroeconomic features, LightGBM classification and return-prediction models, and SHAP-based explanations. built an interactive dashboard for stock research, portfolio analysis, market rankings, technical indicators, and Gemini-generated natural-language analysis.

trained a 60-million-parameter T5 model on 30,000 English-to-Spanish and English-to-French translation pairs. then built gradient attribution-patching tools with custom PyTorch hooks to study how language information moves through six cross-attention layers and 48 attention heads. contrastive benchmarks and token-level heatmaps helped distinguish shared semantic representations from language-specific generation behavior.

implemented core operating system components including thread scheduling, user programs, virtual memory, and file systems. developed system calls, process management, and memory allocation algorithms while working with C programming and low-level system design to build a functional operating system kernel.

developed a machine learning solution to predict customer churn from unstructured and incomplete CRM data, implementing preprocessing and engineering 15+ predictive features such as account age and login frequency, achieving 96% accuracy with gradient boosting (LightGBM). engineered a data-driven classification system to evaluate startup success using custom business logic on operational status, years active, and funding levels, creating robust labels for supervised learning. built and tuned XGBoost models with custom cross-validation, resampling, and advanced preprocessing including target and one-hot encoding, optimizing F1 scores via manual grid search.

Worked with Dr. Karthik Jagadeesh to streamline GWAS and brain single-cell data analysis using Scanpy, Python3-adapted LDSC, and optimized preprocessing pipelines to identify cell types involved in neurological disorders such as Tourette's. Built a protein similarity latent space using Facebook's ESMFold with UniProt data for feature extraction, enabling advanced analysis of protein relationships and functions.