Your card already knows you.
Make it talk.
Drop in a card statement CSV. offramp reads your habits, makes claims about you, lets you grade those claims — and only then does it earn the right to tell you what to do next.
Meet a fictional New Yorker with six months of very real-looking habits: a ghost gym, four streaming services, and a Trader Joe's dependency.
⛬ parsing and analysis run entirely in this tab — no uploads, no servers, no account. Your statement never leaves the page. The one exception is the optional "second opinion" on the last screen, which you have to switch on yourself.
The mirror
before you see a single number — what do you think you spend?
The mirror
perception vs reality. wealth management is behavioral — this is the behavior.
Things you might not clock
Computed from your own data — no generic benchmarks, just patterns you're inside of.
Set the budget
suggestions = your own median month, not aspirationThe debrief
Monthly burn
Where it goes
Every day, on one canvas
Claims about you
Each one carries the engine's confidence. Some of these it can't actually know from data alone — that's what step 03 is for.
Calibration
Your answers become an eval set. The engine gets scored, then rewires itself.
Eval results
How well does the engine actually know you?
Graded claims
Self-corrections applied
Next best action
The queue
ranked by impact × confidence ÷ effortSecond opinion
the only part of offramp that uses the networkEverything above came from rules I wrote by hand. They find what they were built to find and nothing else. Hand Claude the same summary and it reads the layer above: tensions between what you said you wanted and what you actually do. Optional, opt-in, your own API key.