Experiments
-
Can a coding agent trace where government data goes?
See experiment: Can a coding agent trace where government data goes?Tracing where a government application sends its data is hard to do by hand, so we tested whether an AI coding agent could take on the investigation.
-
Estimating AI acceleration in a project budget
See experiment: Estimating AI acceleration in a project budgetWe built a way for vendors and buyers to see the real effect of AI on a software budget, measured task by task against what past projects actually took, and sharpening as more work gets logged.
-
An accountability trail for AI decisions
See experiment: An accountability trail for AI decisionsWhen an AI shapes a decision about a person, can you later show what happened and who was accountable? We built a thin governance layer that records exactly that.
-
Generative AI as an accessibility testing coach
See experiment: Generative AI as an accessibility testing coachAutomated accessibility tools tell you what failed. We wanted to know if AI could tell you why, and coach a non-expert through fixing it.
-
Design handoff for your AI coding agent
See experiment: Design handoff for your AI coding agentWe built a tool that converts a design screenshot or Figma file into two machine-readable files, giving AI coding agents the vocabulary and rules they need to generate UI that actually matches your brand.
-
Vibe coding platforms in the real world
See experiment: Vibe coding platforms in the real worldWe put Lovable through its paces across a variety of different builds and came away impressed by the speed, skeptical about the depth, and clear on where it breaks down.
-
Approaches to automated name matching
See experiment: Approaches to automated name matchingReconciling duplicate names across disparate datasets is harder than it looks. Here’s what we learned testing algorithmic and AI-based approaches.
-
Crowdsourcing park accessibility info
See experiment: Crowdsourcing park accessibility infoWe tested whether AI tagging and semantic search could replace the periodic manual update cycle that causes government information services to slowly become unreliable.
-
AI powered staff writing assistant
See experiment: AI powered staff writing assistantWe tested whether a language model could reliably check government web copy against an official style guide and give writers specific, actionable feedback without a human editor in the loop.
-
Web apps on embedded devices
See experiment: Web apps on embedded devicesWe figured out how to bundle an entire web application, assets and all, into a single file that runs reliably on a low-power embedded server.