P-01Place evidence
The model doesn't get to pick which source to trust.
Place data is checked through a priority chain of real sources; the model only classifies the final result as open, closed, or uncertain — it never chooses the source itself.
Agentryx AI / Products / M-02In development
Itineva
Travel plans built like websites: versioned, shared, alive.
Generic AI itineraries tend to fail in two specific ways: they invent places that don't exist or have closed, and they ignore travel time until a day's plan turns out to be impossible. Itineva is built to make both structurally hard to happen — every place is checked against real data before it's suggested, and every day is checked against actual travel time before you see the plan. In development; the MVP is complete but not yet released.
Specification
InputsDestination · dates · travelers · taste
OutputVersioned trip page
SharingOne link, always current
StageIn development · MVP complete
01 — Real places, not invented ones
Every place is checked before it's suggested.
Generic AI itineraries invent restaurants that don't exist or recommend places that closed months ago. Itineva checks every place against real business-status data before it makes the list — the model never gets to just guess.
02 — Real time, not guessed time
A day's schedule is checked against how long it actually takes.
After a day is planned, Itineva measures the real travel time between each stop and flags it if the schedule doesn't leave enough room — before you find out the hard way, mid-trip.
03 — A living page, not a dead PDF
Plans change. Pages change with them.
A trip page with versioning keeps every traveler on the same plan. Each version has a stable share link, and the featured version is what the group sees by default.
Proof
P-02Route evidence
Travel time is measured, not estimated by a model.
Real routing data checks the time between each stop on a day's plan after it's drafted, and flags any day where the schedule doesn't actually hold up.
P-03Correction evidence
A test that flattered our own approach got re-checked, and reversed.
An internal comparison first showed a simpler approach beating Itineva's own pipeline — then a second look found the test itself was biased (a cheaper config, a shared method both sides relied on). Re-run fairly, the result reversed. The methodology got fixed before the product did.
“Real places, real time — not a guess dressed up as a plan.”