Pathfinder reads your real Gemini usage, replays it against the 3.x models inside your own project, and tells you exactly where every workload should land — and what it will actually cost. Measured, not guessed.
Every stage — reading logs, de-identifying with Cloud DLP, replaying traffic on Vertex AI — executes inside your own Google Cloud project, in the region you choose. No prompt, response, or bill is ever copied to Tilicho or any third party. You run it on your infrastructure; you keep every byte.
It opens with a Survey — is the project even ready, and what will it cost? — then five stages each hand the next a cleaner artifact. ● marks a stage where a human decides — pick any to see what it does.
List prices are per token — but your bill is that price times how many tokens a model actually generates, and that count swings wildly from one workload to the next. So a 5× price sheet tells you almost nothing about the real cost: some workloads get cheaper, some get pricier. That's why Pathfinder replays your own traffic instead of doing spreadsheet math. Three kinds of things it turns up:
A model can list at five times the price per token and still cost less on one workload — and more on the next. What you pay is price × the number of tokens a model generates, and those counts vary hugely from one task to another. There's no shortcut from the price sheet to the bill — you have to measure your own traffic to know.
Before you migrate anything, the analysis reads how your current setup behaves — and routinely turns up workloads that are already misconfigured: answers cut off mid-response, malformed output, budget quietly wasted. Worth fixing on their own, migration or not.
Sometimes the model a workload needs isn't available in your region, or there isn't enough data to be sure. Instead of guessing, Pathfinder labels every figure by how it was obtained — measured, estimated, or assumed — and routes the uncertain ones to a person. You always know what's solid and what isn't.
Every cost is labelled with how we got it — and a total is only as reliable as the weakest number inside it.
So you can always tell what to trust. Was a figure actually measured on your traffic, or is it just a default we assumed? Four labels, strongest evidence to weakest:
The clock is the only thing not negotiable.