Project Portfolio

Led technical architecture for a platform that automatically designs fixed-wing drones — selecting optimal components from catalogs and tuning airframe geometry to meet user-specified mission requirements. Powered by a custom evolutionary search algorithm (genetic algorithm); deployed on AWS cloud infrastructure managed as code; coordinated a distributed engineering team across multiple repositories.

Transformed a 3,000-line research script into a production system estimating how many machine learning experts work at each of 400+ organizations, inferred from 250,000+ employee profiles. Redesigned the statistical model from scratch to correctly account for correlated AI classifiers and produce honest, statistically valid confidence intervals.

A mobile-friendly web app that works without internet connectivity (an "offline-first" Progressive Web App) for waste management teams at large events such as music festivals. Features real-time multi-user data sync, interactive event maps, automated client-facing report generation via Google Slides and Sheets, and cloud infrastructure managed as code.

Mathematical framework for allocating advertising budgets across 18+ channels to maximize campaign reach. Built a statistical model for how frequently people engage with each channel, with a probabilistic correction for imperfect demographic targeting, and formulated the whole problem as a mathematical optimization. Documented in a 14-page specification derived from first principles.

3D computer modeling of eggplants in a 30-meter solar greenhouse with 4,032 plants. Simulated sunlight distribution through the canopy (Monte Carlo ray tracing), used Bayesian optimization to tune 8 architectural parameters, and applied numerical integration to average results over the daily solar cycle — achieving a 29% improvement in photosynthesis efficiency.

Publication-quality figures for a connectomics (brain-wiring mapping) and brain simulation research team. Created multi-scale charts spanning many orders of magnitude, 3D visualizations of computational requirements, and custom heatmaps of neural recording coverage — across four iterative phases through final publication-ready deliverables.

High-performance portfolio rebalancing library written in Rust, using a convex optimization solver (Clarabel.rs). Extended the classical Markowitz mean-variance framework — the standard mathematical basis for balancing risk and return — with novel soft risk constraints; callable from Python. Achieved a 50× speed improvement over the original Python implementation (from ~5 ms to ~100 μs per solve).

Contributed to Kraken's first IRS-compliant tax reporting service. Built in Rust with end-to-end encrypted data exchange with external tax vendors, server monitoring and alerting, and cloud infrastructure managed as code. Served approximately 4,000 users in the first weeks of deployment.

Data pipeline to transform large, irregularly spaced seafloor depth measurements (2GB+) into uniform high-resolution maps (GeoTIFFs at 10m resolution) for GIS analysis. Diagnosed and fixed a memory bottleneck that would have required 769GB of RAM; applied data thinning, trend removal, and spatial boundary masking for correct geographic alignment.

Implemented a graph-based matching algorithm from a research paper that rearranges a matrix's rows and columns to improve its numerical properties for downstream computations. Handles values spanning 37 orders of magnitude. Validated against MATLAB's built-in reference implementation; client awarded a bonus doubling the contracted amount.