Research Agenda

The Web Framework AI Can Actually Understand

A Python web framework designed for AI-first development. Structured so LLMs can read, modify, and extend your code reliably. Runs on a $2.50 VPS or a Raspberry Pi. 16KB JavaScript footprint. Built on FastAPI, Pydantic, Starlette. MIT licensed.

Project Templates That Learn From Usage

Project scaffolding for any language. 200 million downloads. Now we're modernizing it to make it a core dependency of Air. The next phase: AI that tracks how projects evolve from templates, detects patterns across thousands of codebases, and lets improvements flow both directions. Templates get smarter because people use them.

Commodity Hardware Matches Cloud Performance at 10x Lower Cost

AI inference is becoming essential infrastructure, but costs assume US/EU pricing and bandwidth. A developer in Manila or Lagos shouldn't pay more for worse access. The bottlenecks are known. The optimizations exist. We're building the stack that makes $2 inference real.

Vibe Coding Works in Any Language

AI can now turn natural language into working software, but current tools charge US prices and only work in English. A farmer in Mindanao describing an inventory tracker. A small business owner in Brazil explaining a booking system. A teacher in rural India sketching classroom tools. They describe what they need in their own language. The software gets built.

Nanotech Pathways to Sub-3nm Fabrication

The hardware gap is widening. Developers in wealthy countries run models locally. Everyone else waits for API calls over slow connections. The path forward is smaller, cheaper, more efficient silicon. We're developing novel fabrication approaches at international facilities where frontier research is possible without billion-dollar fabs.

Crowdsourced Human Decisions Train Continent-Scale Grid AI

Deep reinforcement learning for power grid optimization, delivered through a strategy game. RTE France proved RL works for grid control with Grid2Op, but on tiny 14-118 node research grids. We want to scale to continent-size networks (20,000+ substations) and generate training data from millions of human decisions. The game is the data collection mechanism. The AI is the research output.