Posts

FlowNode: guesses are cheap, checks are cheaper

The problem. My partner from True Bots is building AI-enabled engineering simulation software. The existing simulation software is very expensive and hard to use, even though the formulas behind it are public. I took the AI subsystems and the chat, with a focus on running local models, which are limited by the hardware they run on.

What I built. A query's goal is a physically accurate thermal system. So I designed the loop so that guesses come easy and checking them is inexpensive: the harness iterates until it produces a reasonable system that the app can diagram and simulate. That lets a weaker local model reach a correct answer by taking more turns.

Around it: model suggestions that fit the machine's video memory, adaptive context, cache reuse between turns, local-first chat history, structured-output adapters for local and cloud models, and Model Lab, an in-app benchmark that runs models through a gauntlet and grades them against known answers on physical invariants.

What happened. 32 commits and about 38,000 lines between June and September 2026. The product is in development.

What I learned. Local AI is limited by its hardware, so design around the limit: make guesses easy, make checking them inexpensive, and let the harness iterate until the answer holds up.