I'm a Lead Machine Learning Engineer at Dstillery. Most of my work there is data engineering, MLOps, and the pipelines behind our models. I have a master's in computer science from UMass Amherst, with a concentration in data science.

Outside of work I make games for fun, and I've released around 20 of them at unitedfailures.itch.io.

• burning-wood-versus is an online arena game I wrote in Rust over a summer. I built the multiplayer netcode, with an authoritative server, client prediction, and delta updates.
• its-caturday runs GPT-2 and RoBERTa locally inside a Unity game through ONNX Runtime. I wrote the inference and the tokenizers in C#. GPT-2 suggests and autocompletes words as you type, and RoBERTa scores how well each tweet fits the trending topic.
• card-tree-puzzle is a deckbuilding prototype played on a randomly generated tree. The game logic is plain C# with no Unity dependency, and it's unit tested.
• diggity-diggity-dash is a mole racing game where you dig through terrain built on marching squares. The terrain is split into chunks, and digging only rebuilds the ones it touches. The AI racers use A* to find their way as the tunnels change.
• pikmin-squad-prototype is a Pikmin inspired prototype built on an entity component system I wrote on top of Unity.
• say-cheese is a game about spotting suspects in a crowd through security cameras. I built the procedural crowds, where everyone is put together from layered sprites and the suspects hide among decoys.
• courier-crusaders is a management game about running a fantasy courier company. I wrote the probability model behind events on the road, which shows your party's exact odds before you send them out.

• DISAPERE is a NAACL 2022 paper I co-authored as a research assistant in the Information Extraction and Synthesis Lab. It's a dataset of peer review discussions annotated with their discourse structure. I built early BERT baselines on the review data.
• implicit_reward_optimization is reinforcement learning research I did with a PhD student in the Autonomous Learning Laboratory. It learns a reward function and a discount function that keep an agent on the real goal. That lets hand-made helper rewards speed up learning without leading it astray when they're wrong. I wrote it in PyTorch and ran the hyperparameter sweeps on a Slurm cluster with Hydra and Nevergrad.
• GANDataGeneration is a paper I co-wrote on training classifiers with synthetic data from GANs, to anonymize a dataset or to balance one. I wrote most of the GAN code, including the conditional convolutional GAN that worked best.
• Image2Minecraft is a group project that turns a single image into a Minecraft build. I wrote the voxelizer, the matching of block textures to each voxel, and a CMA-ES search over block choices that scores each attempt by comparing renders.
• Subreddit-Stock-Prediction was a group project for my NLP class on whether sentiment in subreddits like wallstreetbets tracks the market. I built the pipeline that pulls seven years of posts and comments from Reddit and labels them with stock prices. I also turned the sentiment of each post and its comments into features for a PyTorch classifier that predicts whether the stock a post mentions rises over the following week.
• rl-policy-improvement was a project for my reinforcement learning class. It searches for new policies with CMA-ES and keeps the ones that beat the old policy with high confidence, using importance sampling bounds on the old policy's data.

• I've had fixes merged into the Turbo game SDK, like a bounds check on network reads and tinting for nine slice sprites.
• For burning-wood-versus I added 2D point lights to my fork of the Turbo engine, from the renderer and runtime bindings up through the SDK.