Projects

Interactive mental-health NLP stack: intent classification and domain-specific responses using PyTorch, Hugging Face Transformers, and TensorFlow/Keras. Fine-tuned LLaMA 3.2 1B with LoRA adapters (including efficient QLoRA via Unsloth) for clinical text generation, alongside BERT-style classifiers for supervised mental-health labeling. Reached 0.79 Macro-F1 on aggregated clinical data while cutting training VRAM by 50%+; built an ETL pipeline over 53,000+ text samples for fine-tuning.

SeeMyRace is a full-stack, Dockerized Flask platform for race photography—Google Drive ingest, bib and selfie search, and Google OAuth. I led bib-number detection research (including SWT), shipped much of the main site UI, flagging and moderation flows, and backend connectors into the ML inference pipeline. For OCR, I evaluated Tesseract and PaddleOCR; PaddleOCR looked strongest for our task, but at the time its docs skewed Chinese-heavy and were hard to operationalize for a lean MVP.

I architected DeepFace-based semantic facial recognition targeting sub-50ms inference and an asynchronous ingestion pipeline that kept heavy model work off the request path, with UI staying responsive (under 100ms) during bulk uploads. Separately, the product’s face matching also relies on PostgreSQL with pgvector (cosine similarity over embeddings)—that vector search path wasn’t my direct implementation focus. The repo has since gained contributors and more CV plumbing (e.g. YOLOv8, InsightFace), while goals stay aligned.

A Princeton University Studio Lab research initiative to create the largest unified dataset for Indian political linguistic analysis. Engineered a resumable Selenium crawler to index the Parliament Digital Library (handling dynamic pagination and session state), processed 40+ GB of election speech audio (Modi & Gandhi, 2014/2019 campaigns) through an AWS Transcribe/Translate pipeline, and digitized 40+ years (1981–2024) of Lok Sabha and Rajya Sabha debates. Designed a PyMuPDF and FuzzyWuzzy text extraction engine to structure thousands of raw PDF statements, mapping OCR output to standardized Ministry entities at scale.

Engineered a TrueNAS Scale home server cluster with a 24TB storage array. Orchestrated containerized microservices (Jellyfin, Immich, Tdarr) for high availability and seamless media transcoding, secured via Cloudflare Tunnels and Tailscale VPN.

HateBlocker pairs a Chrome extension (content scripts, JavaScript/HTML/CSS) with a FastAPI backend to score text on web pages and hide hateful or offensive content—similar to an ad blocker for toxic language. Trained on the Davidson Hate Speech and Offensive Language dataset (~24,800 labeled tweets): stacked 5,000-dimensional TF-IDF with linguistic metadata (tweet length, punctuation, hashtags). BERT sequence classification reached 91.5% accuracy; XGBoost reached 90.72%, with SVM, logistic regression, random forest, and naive Bayes as baselines for a three-way hateful / offensive / neither task.

Diabetes risk modeling on BRFSS 2015 tabular survey data (400,000+ U.S. respondents). Compared SVM, decision trees, random forests, logistic regression, and Keras Tuner-driven neural networks, with explicit attention to class imbalance and the recall–precision tradeoff for screening. Reported tuned accuracies around 75.3% (logistic regression), 74.8% (SVM), and 74.3% (decision tree); GenHealth, BMI, age, high blood pressure, and income surfaced as consistently important risk factors across models.

Designed a multimodal predictive regression pipeline fusing dense text embeddings (sentence-transformers/all-mpnet-base-v2) with 100+ prosodic audio features (pitch variance, speech rate, pause density). Used Cosine Similarity over multi-turn dialogue structures to model conversational coherence. Deployed K-Fold Cross-Validation (k=5) to minimize prediction error, achieving MAE < 0.4 on normalized score predictions and Pearson r > 0.70 correlation with human rater scores. Applied SHAP explainability to identify which linguistic and tonal features most strongly predicted candidate ratings across competency dimensions.

Architected a secure, containerized full-stack marketplace application for CU Boulder students using Node.js/Express and PostgreSQL. Implemented bcrypt-hashed authentication with session-based authorization, RESTful API endpoints with async/await controller logic, and a Dockerized deployment with isolated service containers. Built persistent user records with real-time cart state management, supporting concurrent sessions and transactional integrity for listings, orders, and user profiles.