Writing and building around the intersection of capital and AI capability curves.
AI is systematically undermeasured across the economy. Most evaluations, environments, and the leverage inside companies are still uncertain. I work alongside that with the conviction that coding agents are general-purpose agents that transform the economy.
The thesis: every operating segment in the economy is being re-priced by AI, segment by segment, and the market is slow to recognize it. I build the low-level substrates and the evals that sit on top of them: a CLI for agent-authored OpenUSD, a capability-to-eval feed, and an RL environment for project controls. Each is written up below.
A CLI authoring tool that lets agents create OpenUSD scenes through typed JSON operations instead of brittle prose edits. Better agent ergonomics and training-data acquisition for 3-D is critical to the future of 3-D pipelines, sim2real, and the design of the built world. usdq enables agents to author, validate, inspect, repair, and render.
A construction-style agent environment for schedules, critical path, float, XER/XML, and delay review, turning controls tasks into inspectable, scorable episodes with auditable state. The demo is live at projectcontrols.elidukes.xyz, go run an agent on a real schedule.
Cataloguing and analyzing AI as it intersects with traditional vertical software, new services, and new paradigms altogether.
Maps a job or workflow to public eval coverage, adjacent capability evidence, and active human-data demand, separating direct eval signal from proxy, so a workflow that only maps to generic document QA isn't sold as solved.