Making AI faster, smarter,
and more accessible.

A full-stack RAG system that quantifies how much you can trust an AI's answer. Documents are ingested, vectorized, and run through a multi-faceted confidence engine — combining consistency, sensitivity, and novelty scores — to produce a single explainable reliability score per query. Results stream to a React frontend; all activity is logged in PostgreSQL. This project was developed as a Virginia Tech Capstone, with the problem statement provided by Boeing.

An AI-powered personal productivity app generator built at HooHacks 2026. Describe any app you want — a habit tracker, spending dashboard, mood journal — and Mugen builds it instantly as a fully functional React app, pre-filled with your own real data. Paste in raw text and Claude infers your schema, links your data sources, and ships a complete app in seconds.

A fine-tuned medical chatbot built on Qwen-8B, designed to simplify complex medical jargon for older patients. Fine-tuned using SFT and LoRA on 20,000 doctor-patient dialogues, accelerated with vLLM and AWQ quantization, and evaluated with a 48.5 SARI score and a 4.5/5 usability rating.

A two-tier cascade pipeline that classifies depression severity in social media posts across four PHQ-9-aligned levels. A recall-optimised DistilBERT sentinel gates posts to a QLoRA fine-tuned Llama 3.1-8B reasoning engine, achieving 100% severe recall and strong macro F1 — while keeping inference efficient and safety-first.

An NLP interpretability suite that probes whether a fine-tuned RoBERTa-large model actually reasons the way it claims to. Two custom metrics — Latent Alignment Score (LAS) and Causal Sensitivity Index (CSI) — measure how closely the model's internal state matches human-grounded rationales vs. post-hoc justification on the adversarial ANLI benchmark.

Real-time raccoon and rodent detection using lightweight YOLO models, deployed on edge hardware. Trained and compared YOLOv8n, YOLOv11n, and YOLOv11m across two custom datasets, stress-tested with synthetic augmentations, and deployed to a Raspberry Pi and desktop via a React dashboard.

A full-stack mobile app that scores used car listings to surface real deals. Type in a car's details in plain text and get a data-driven value score back. This mechanism is powered by a neural network, an NLP parsing pipeline, and a live database fed by a web scraper.

A top-down wave shooter built in Python and Pygame ported to HTML+JS. Dodge and shoot endless waves of escalating ghosts, collect 8 powerups, and survive as long as possible. Playable in the browser.