Engine
The intent engine behind everything we build.
A technical look at how Camino turns fragmented signals into a persistent, activatable understanding of intent - across conversations, content, and the open web.
How it works
Camino isn't a single model or a single score. It's a four-stage pipeline that turns raw, fragmented signal into a structured understanding of intent - and keeps that understanding current as people keep interacting.
Capture. Every surface we touch - a conversational turn inside a Camino Apps experience, a scroll or dwell event, a piece of content itself, a signal an advertiser already holds - is normalized into a common event schema. Nothing is useful until it's comparable.
Model. Transformer-based embeddings map that normalized signal into a shared intent space. Rather than collapsing a person into a single propensity score, we build a persistent, per-user intent graph: a structured, evolving map of what someone cares about, how strongly, and why - refreshed continuously rather than recomputed in nightly batches.
Activate. A low-latency inference layer serves that graph wherever it's needed - shaping a response inside a Camino Apps conversation in milliseconds, riding along as a bid-time signal in Camino for Publishers' monetization platform, or landing as structured audience data inside a publisher's CRM and Google's Publisher Provided Signals (PPS) via our audience intelligence product.
Learn. Every activation is itself a new signal. Outcomes feed back into the graph, so the engine recalibrates continuously - it gets more accurate the more it's used, not just the more it's retrained.
Design principles
Identity without surveillance.
Signal capture is first-party and consent-based. The graph is built from what people tell us and do on our surfaces - not from cross-site tracking.
Composable by design.
The same intent graph powers a publishing experience, a bidstream integration and a CRM export. One representation of intent, activated through different APIs.
Low latency, high fidelity.
Inference is built to run inside the ad call and inside the conversation - not as an offline enrichment step that arrives too late to matter.
Legible, not opaque.
Outputs are structured and interpretable. Partners see the intent signals behind a decision, not just a number they have to trust.
Where it shows up
One engine, two products.