Technology
Rivier is three things working as one: a memory you own, a learned AI that reasons on top of it, and the rails that let it act and pay on your behalf. Each layer is built, not borrowed — and each one compounds the moment you use it.
The stack
Most assistants own the chat and rent everything else. Rivier inverts that: it owns the layers that matter — your memory, the reasoning, and the means to act — and rides whatever chat surface you prefer.
Everything Rivier knows about you — taste, plans, the way you like things done — lives in an encrypted personal graph. You own it. It moves with you. No assistant gets to keep it.
It runs on Tenzro, a decentralized network for identity, memory, compute and payments. Assistants receive the personalisation; they never touch the raw graph behind it.
Rivier brings its own intelligence. Rivier One is a learned conductor that routes every request, in the moment, to the model that answers it best per dollar — a cheap specialist we own, an open model on the network, or a frontier model when the task earns it.
And it checks its own work: it knows when an answer is good enough and when to reach for a stronger model. You get frontier-quality results, without paying frontier prices on the everyday work.
Rivier doesn't stop at an answer — it does the thing. A scoped agent identity and a settlement spine let it book, buy and pay on your behalf, inside the limits you set, on your explicit yes every time.
RIVR is the accounting layer: one balance, seen in your own currency, with every confirmed action carrying a signed receipt you can audit.
And it settles for real — on Canton, the same institutional-grade ledger the markets run on. Your balance isn't a number in an app. It's money that actually clears.
Rivier One
The industry split intelligence and compute into separate camps. Rivier One sits in the gap: a learned conductor that routes across decentralized and frontier compute alike, choosing per request which world a task belongs in — optimised for quality per dollar, with cost, latency and privacy as first-class inputs.
Own
Narrow models we own, tuned to the tasks our users actually run — frequent work at near-zero marginal cost.
Network
An open market of providers on the Tenzro network, always at the lowest going rate — cheaper than any single provider can offer.
Frontier
The strongest frontier APIs, called precisely when a task earns it — never categorically off the table, never the default.
Why it widens with use
Rivier One learns continuously. Each request writes a trace — the task, what ran, what it cost, whether it landed — to a ledger that retrains the system on a rolling schedule. The product gets cheaper and sharper the more it is used.
A request is routed, run, and verified — answered at the right tier for the job.
The full trace — plan, model, cost, outcome — is written to the telemetry ledger.
The conductor, the verifier and the owned specialists retrain on that ledger.
Routing gets sharper, escalation gets rarer, cost drops — and the gap compounds.
The moat
Architecture can be cloned. The accumulating record of how real people actually use Rivier — the routing traces, the outcomes, the task clusters worth building a specialist for — cannot. It is the realised distribution of our own demand, and it feeds three retraining loops at once.
So the system that improves is not just one model. It's the conductor, the verifier that gates cost, and the owned specialists that absorb the most frequent work — all sharpening on traffic no one else has. The lead widens with volume, not just time.
For the engineer
Rivier is invite-only right now. Leave your email and we'll reach out when there's a place for you.
Request early access