tablesalt
Drop a CSV.
Ask a question.
See generative UI.
Drop a CSV. The agent profiles your data, picks the right kind of answer for each question — a chart, a stat card, a table — and renders it with motion that exists to communicate state, not to decorate. No upload, no backend, no signup. Everything runs in your browser.
I built this to show four things in one place: that an AI agent can produce useful product UI, not just text; that its accuracy can be measured with a real eval harness running live (scroll down and press the button); that streaming UIs can feel intentional instead of janky (powered by streamfield, a primitive I extracted and published to npm); and that the whole thing can ship with zero backend.
The eval scoreboard.
Most AI-for-data demos hide their accuracy. Click the button and 12 hand-labeled questions run against the live model now. Each case is scored on three things — does it pick the right rendering, does the SQL execute, and does the SQL return the same answer as the reference query.