FastCombo — Deep RL for Combinatorial Auctions

Recurrent PPO for partially observed CATS auctions with direct item pricing

Collaborating with Bryant faculty and a graduate student on a recurrent PPO actor that reads public histories from partially observed, single-minded CATS combinatorial auctions and directly emits one item-price vector. On Paths, it reached 92.4% deterministic exact stage-1 clearing, above the published 88% Bayesian Paths result under a different protocol.

Python PyTorch Recurrent PPO Partially Observed Auctions Direct Item Pricing Game Theory

Ask Tupper — Hardened Campus Chatbot

QLoRA fine-tune of Qwen3-32B with layered prompt-injection defense

Security-hardened campus assistant replacing a production system found vulnerable to prompt injection. Fine-tuned Qwen3-32B (32.9B parameters) with QLoRA on a single consumer GPU in 90 minutes, then defended by a six-layer injection-protection pipeline that drops attack success from 20% to 0% across 15 tested vectors. A custom RoBERTa classifier provides 95.3% injection recall, and hybrid RAG over 530 scraped university documents lifts factual accuracy from 60% to 74%.

Python PyTorch QLoRA vLLM RoBERTa RAG FastAPI Docker
View repository →

RAPTOR: Route-Audited Patrol Territory Optimization

Routing-aware graph partitioning with integer staffing and audit gates

Submitted ALENEX 2027 paper introducing an offline route-oracle-audited optimizer for patrol territory design. RAPTOR searches partition-staffing candidates, emits route-audited district geometries and officer-per-district recommendations, and reports checksum-backed evidence across Providence, Boston, and Seattle street-network instances while keeping deployment, legal, and causal claims out of scope.

Python Graph Partitioning Integer Staffing Graph Algorithms GeoPandas OSMnx

GuidePost — Patrol Routing and Decision Support

Authored research with paired retrospective evaluation

Across 264 paired Boston conditions with identical peaks and ordering, KDE-penalized routes increased retrospective holdout-incident encounters by 26.3 and route coverage by 16.7 percentage points versus shortest-path routes. The work supports a patrol decision-support platform for fleet coordination, dispatch recovery, command review, and authenticated, audit-logged workflows across web, iOS, and Android.

Graph Routing Paired Evaluation Decision Support Web iOS Android

Responsible ML and LLM-Assisted Insurance Review

Model governance, reliability review, and LLM workflow adoption

Contributing to ML model governance through bias assessment, mitigation analysis, reliability review, and production-readiness evaluation using Python, multivariate methods, and AWS SageMaker. Developing LLM-assisted image-analysis workflows that compare visual hazard descriptions with inspection reports to surface discrepancies for underwriting review, while supporting teamwide adoption of LLM workflows and pipelines through hands-on coding and workflow integration.

Python Statistical Analysis Multivariate Methods Bias Assessment AWS SageMaker LLM Workflows

SafeWalk — Route Comparison with Public Incident Data

MIT-licensed iOS, backend, and reproducible multicity research system

Open-source system that helps users compare shortest and risk-weighted walking routes. Its full stack combines SwiftUI, Fastify/FastAPI services, igraph routing, time-binned KDE surfaces, and reproducible public-data pipelines. Multicity evaluations with strict temporal holdout compared KDE, XGBoost, and graph neural networks; KDE achieved at least 50% greater corridor-based holdout-incident avoidance in every city.

SwiftUI Fastify FastAPI igraph KDE XGBoost Graph Neural Networks
View repository →

Machine Learning Internship

Forecasting, classification, semantic search, and segmentation

Built production ML systems across the analytics stack: time-series forecasting models with automated feature engineering and scenario simulation for inventory and revenue planning; an image classification pipeline reaching 96–99% accuracy across hierarchical product taxonomies with automated metadata extraction and label inconsistency detection; and a semantic search system using vector embeddings to translate natural-language business questions into executable SQL over product and customer data.

Python Scikit-learn Pandas CNN OpenAI API Gemini API UMAP Embeddings

Neural Networks from Scratch

Autograd engine and language modeling from first principles

Implemented a scalar-level automatic differentiation engine with topological-sort backpropagation, built multi-layer perceptrons from first principles, and trained character-level language models on 32K names with learned embeddings and cross-entropy optimization.

Python Autograd Backpropagation Language Models

Interactive ML Visualizations

Browser-based tools for ML internals

Browser-based tools for understanding machine learning internals: a neural network playground with real-time decision boundary visualization, a 3D gradient descent visualizer comparing SGD, Momentum, and Adam optimizers, and a convolutional filter explorer demonstrating kernel operations with live input/output rendering.

JavaScript Three.js Canvas Visualization
Try demos →