Things I built

The Air Force has a problem called the Airlift Logistics Challenge: a Pickup-and-Delivery instance where the number of aircraft, cargo items, and airfields all change between problem instances, and a policy model has to output a variable number of actions at each step. No standard neural architecture handles reasoning across each modality elegantly.

So I designed Flexible Input Networks: a class of architectures that learn over "sets of sets" of arbitrary cardinality and output variable-length action sets in a single forward pass. This architecture requires no autoregression, no recurrence, and no padding tricks. The formalism generalizes beyond logistics to any multi-modal, variable-cardinality problem.

A summer at the NSA's math group, trying to answer: can you make an LLM hallucinate less on technical text by bolting an expert-curated ontology into its encoder, rather than by making it bigger? We built an attention-based autoencoder whose encoder was fused with a domain knowledge base, then evaluated it on semantic extraction over scientifically nuanced documents.

Russia destroyed a satellite with a kinetic ASAT in late 2021 and left a debris field that's still in orbit. In response, NATO commissioned a series of space deterrence wargames meant to develop a framework spanning the diplomatic, informational, military, and economic instruments of power. I led a 29-cadet research team that helped to construct the inagural wargame.

The output was a set of policy frameworks for national and alliance authorities detailing how the alliance should signal, escalate, and respond to attacks in a domain where attribution is hard, collateral damage is quasi-permanent, and there are few established rules.

This system is agentic malware that gets into your email account, reads your existing threads, and replies inside those threads with spear-phishing messages that sound like you wrote them, threading into a real conversation already in your inbox. Imagine you ask a friend for help on a project. Our malware responds to you in that thread from your friend's account with a phishing link that downloads a trojan. The malware replicates and repeats.

The point wasn't to ship an attack. The point was to show what the social-engineering threat surface looks like the moment LLMs become competent at reading context and writing convincingly. The result is uncomfortably good. The defensive posture most organizations have today — "users will spot phishing if it sounds off" — is now a bet against a model that has read your last six months of email.

A P300 speller is a system that lets you type by thinking. Letters flash on a grid; when the one you're attending to flashes, your brain produces a measurable spike about 300ms later. A good classifier turns those spikes into characters.

I built the whole pipeline end-to-end including the EEG helmet hardware, the flashing stimulus GUI, the data collection harness, and a recurrent CNN classifier that translates raw voltages into typed text. The primary use case is communication for people with ALS or other severe motor disabilities, who have intact cognition but no reliable way to output language.

The class assignment: train a vision-only autonomous rover, then race it head-to-head against the other students' rovers on a track none of us had seen using only a camera.

I started with behavioral cloning from a handful of human-driven laps to get a working baseline, then fine-tuned with RL, rewarding faster completion of subsections and punishing off-track behavior.

Show the system one photo of a person, and it pulls every other photo of that person out of a 40,000+ sports-photo database. The backbone is a CNN transfer-learned from a face-recognition base model and fine-tuned with a triplet loss so that embeddings of the same person cluster together, and different people drift apart. Queries hit a vector database with cosine similarity over the learned embedding space.