Colman Humphrey

• Developed an optimization algorithm in Python & Rust to generate multi-axis machining adjustments from part measurements. Integrated geometric analysis, statistical modeling and adaptive feature grouping to maximize conformance in subsequent runs while minimizing correction complexity.
• Engineered a high-performance metrology data extractor in C#, cutting processing time from an hour to under two minutes per part; enhanced flexibility for diverse part types while maintaining geometric precision.
• Architected and implemented a high-precision tolerance solver, integrating complex GD&T principles; developed expertise in tolerancing and datum reference frames to resolve intricate inspection challenges across Formlogic.
• Reverse-engineered industry-standard metrology software calculations through rigorous testing and documentation analysis; identified and resolved discrepancies to ensure accurate replication of standard reports.

Slight: A central data dispatcher that builds data apps from parameterized SQL, providing intuitive interfaces for all stakeholders to access data how they want it. Spreadsheets, APIs, the website UI, Python, anywhere: anyone who can write SQL can generate a robust REST API backed by their data warehouse.

Built a functional product, secured initial funding, and established an early customer base in a challenging market. Unfortunately we did not gain enough traction to raise our seed round. A demo video of Slight is available .

• Architected and developed Slight’s core platform using Rust for high-performance backend operations (complemented with Go components) and React (Next.js) for the responsive frontend.
• Led successful pre-seed fundraising, securing $1.1 million to accelerate product development and market entry.
• Recruited and led a team of five engineers through the development and launch of Slight’s initial public release and SaaS-based commercial platform. Guided product strategy and maintained hands-on involvement, resulting in the onboarding of 16 companies.
• Engineered Slight’s Google Sheets integration, enabling seamless execution of parameterized SQL queries directly within spreadsheets, significantly enhancing data accessibility for non-technical users.

• Led and executed the technical development of a suite of tools for shift, driver, and vehicle assignment, crucial to Via’s expansion into new B2B operations. Recruited and managed a team of three, balancing hands-on coding and architecture design with project management, ensuring code quality, timely delivery, and team mentorship.
• Implemented and deployed serverless applications using Rust and Python on AWS Lambda, forming the core of the tool suite. Integrated advanced integer programming optimization and led the development of a React-based web UI, enhancing operations teams’ efficiency for daily and weekly team workflows.
• Pioneered an algorithm for scheduling in-advance ride requests, optimizing supply efficiency by smoothing demand. Coordinated with the engineering team to integrate within Via’s existing infrastructure, resulting in near-elimination of unserved rides.

• Developed an internal R package to streamline data flow between data warehouses, AWS S3, R, and local files, facilitating merges/upserts, asynchronous queries, and CLI tool usage.
• Led a change in app behavior concerning walking and detours, based on retention analyses using random effects models and boosted tree models via XGBoost. Additionally, developed a real-time NLP pipeline to reroute high-priority unread customer messages.
• Optimized operational reporting by automating 5+ hour weekly data processes into on-demand Slack bot tools running in minutes. Transformed reactive problem-solving for ops teams into daily proactive issue prevention.

• Predicted TV ad viewing of users using YouTube cookies to optimize ad delivery.
• Peer bonus for internal consulting: automated estimation using hypergeometric distribution.

Dissertation:

Matching: The Search for Control

Advisors:

Prof. Shane Jensen, Prof. Dylan Small

Summary:

Developed novel approach for evaluating matches in causal inference. Used machine learning to assess match quality: poor matches are identified when treated units are highly predictable. This method provides insight into match quality, improving reliability and interpretability of causal estimates.

Gold medal, Trinity Scholar, Hamilton Prize

Selected Published Papers

• Using Predictability to Improve Matching of Urban Locations in Philadelphia. Annals of Applied Statistics
• Modeling lottery incentives for daily adherence. Statistics in Medicine
• A Tale of Two Twitterspheres: Political Microblogging During and After the 2016 Primary and Presidential Debates. Journal of Marketing Research
• Urban vibrancy and safety in Philadelphia. Environment and Planning B

R Packages Developed

• glmmboot: for easy bootstrapping with random effects.
• matchfinder: for creating and evaluating predictability-based matched sets.