I build AI systems that see the world from above. Satellite imagery, aerial photogrammetry, and lidar are my native data formats.
I work where aerial imagery meets machine learning.
I co-founded Lightspace Labs, an AI-powered competitive intelligence SaaS built on Claude. By day I work as a photogrammetrist at ESP Associates, building AI-assisted automations that wrap legacy survey software.
I hold an M.S. in Global Health from Duke, where my thesis trained ConvNeXt and DeepLabV3+ models on PlanetScope satellite imagery to detect coal-plant emissions in Ulaanbaatar. Before Duke I studied marine biology at UCLA and rowed Division I.
Lately: training models on overhead imagery, wrapping outdated tools in AI, and shipping small things on weekends.
A working log of what I have built in AI and machine learning. Live products, research, and experiments. Each entry is dated and linked where available.
AI competitive intelligence. Built on Claude.
An AI-powered competitive intelligence SaaS. Five modules per competitor: citation visibility across Perplexity, ChatGPT, and Gemini; marketing strategy extraction; tech-stack detection from public code; pricing intelligence; and a screenshot-driven product deep-dive powered by AI vision. Designed, built, and operated end to end.
A team of AI assistants that runs the back office of a small business.
Nine specialist agents under an orchestrator, each owning one corner of the business: inbox triage, lead capture, review response, SEO, marketing, site monitoring, weekly reporting. Every agent drafts in the background and holds anything customer-facing until the owner approves, with per-agent permission scopes and auto-send unlocked task by task as trust builds. Pivoted from an agent-observability SDK, whose guardrail layer became the permission system.
Satellite imagery becomes draw-verification reports for construction lenders.
A three-agent pipeline. Agent 1 ingests satellite imagery over a bounding box and date range and emits a GeoParquet of H3 hex cells with 20-dimensional feature vectors. Agent 2 classifies. Agent 3 reasons over labeled cells and produces a draw-verification report. Replaces $800 to $2000 human site inspections.
Detecting power-plant emissions from satellite imagery with deep learning.
Trained a two-stage pipeline on 3 m/pixel PlanetScope imagery over Ulaanbaatar, Mongolia. A ConvNeXt Large classifier handles binary plume detection; a DeepLabV3+ model with a ResNet-101 encoder produces pixel-level segmentation. Cross-location evaluation on Bishkek CHP quantifies domain-shift effects without fine-tuning.
Learn aerial photogrammetry the interactive way.
A mobile-first PWA that teaches aerial photogrammetry and GIS from first principles. Every concept pairs with a hands-on demo. Installs to your home screen. Built as a learning project: lessons are written as the material is learned.
Upload a point cloud, get a classification.
Backend plus frontend for classifying LiDAR point clouds. Containerized, deployed on Railway. Handles LAS and LAZ inputs, returns labeled output.
Brand work, side projects, and a discovery directory I built for my sister. Hand-rolled, fast, and easy to maintain.
Find yoga, pilates, and wellness studios across the U.S.
A discovery directory for 2,400+ fitness and wellness studios. Sanity CMS for content, Cloudflare Workers and D1 for the edge API, MindBody integration for live class data. Search by location, browse by discipline, see class schedules.
Brand site for a CMS migration consultancy.
Multi-page marketing site with custom branding, diagonal section dividers, and a clean information architecture for the services on offer.
Independent guides for surveying equipment and GIS gear.
Editorial site comparing surveying instruments with deep explainers on how each one works, plus a catalog of options. Independent. Not a dealer.
Landing page for a wealth advisory firm.
Brand-forward landing page deployed on Cloudflare. Lightweight, no framework, easy for the client to update.
Brand identity and landing page for a logistics firm.
Custom logo work plus a clean single-page site introducing the company and its services.