MS Computer Science @ Northeastern (Khoury) • BS CS & Math, Kansas State
AI engineer focused on security • Looking for Summer 2027 internships & Fall 2027 co-ops

I build AI systems end-to-end, and I keep the AI on a short leash. In Doublethink, the model that holds the secret never reads the web. In HeatSwitch, the LLM reads documents and writes sentences, and every number comes from tested code.

Before grad school I worked in a university SOC, so I tend to ask "how would someone break this?" early. I care about code that's boring to maintain: clear boundaries, good tests, and systems that fail loudly instead of quietly.

Right now: MS CS at Northeastern, building SafeCommit (a safety layer that stops AI agents from taking destructive actions) with NU Launch Labs, and contributing to Hack for LA.

🔐 Doublethink · 🏆 1st Place, Most Startup-able, HACK1984 · Demo An AI research agent that can't leak the confidential data it holds. A local model holds the secret and never reads the web, a cloud model reads the web and never sees the secret, and a fail-closed gate is the only process allowed on the network. Red-teamed with 40 prompt injection attacks: a standard agent leaked the secret in 2, Doublethink in 0. I built the gate, the hash-chained audit ledger, the network guard, the planner, and the orchestration loop. Python Ollama Qwen3 Claude API SQLite

🔥 HeatSwitch · Live app · Demo Tells Boston residents whether a heat pump actually saves them money. Upload a gas, electric, or oil bill, Claude reads it, and pure TypeScript does the math: yearly savings, Mass Save rebates, payback. 144 unit and property-based tests, plus a ranking of 1,249 Boston buildings from the city's BERDO energy data. Built solo at Hackachusetts 2026. Next.js TypeScript Claude API Zod Vitest

🛡️ LLMGuard · Paper · Poster Capstone research on catching tampering with a locally deployed LLM. Once a model is loaded, it's usually trusted blindly. LLMGuard is a layered design that pairs TPM + Linux IMA (load-time file integrity) with TEE-backed behavioral attestation (runtime), so each layer covers the other's blind spot. Includes a threat model, a comparison of five attestation approaches, and a 10-page IEEE-format paper. Security TPM Linux IMA TEE LLMs

📋 JobTrackr · Solo Kanban-style job application tracker with contacts and interview rounds per job. FastAPI + PostgreSQL backend with Alembic migrations, JWT access and refresh tokens, bcrypt password hashing, and per-IP rate limiting on login and signup. React frontend with drag-and-drop, deployed on Vercel and Railway. FastAPI PostgreSQL SQLAlchemy React Tailwind

🌾 AguaCrop · 🏆 1st Place, K-State TAPS Hackathon Irrigation planner that combines three K-State field datasets with 7-day weather forecasts to compute each plot's optimal weekly irrigation, with a React + Chart.js dashboard and emailed PDF plans. Built by a team of 3 in 48 hours. Python Flask React Chart.js

Languages

AI / ML

Backend & Data

Frontend

Security & Tools