The AI stack changes every quarter, and the teams building on it try tools in public long before they buy. DemandCompass reads those signals, scores every account on whether it knows you and whether it is moving, and lets old ones expire.
Teams swap model servers, eval tools and vector stores fast. The Decay Rule ages every signal, so an account that tried you in spring and moved on stops looking hot.
The people starring your repo are not the ones with the budget. Signals roll up from people to the account, so the whole team shows up as one.
Hiring for the AI stack, AI-agent development and AI investment in annual filings say who is spending, and on what, before anyone fills in a form.
Readiness says the account is moving. Awareness says it knows you. Sales-Ready needs both, from more than one source, inside the window.
AI engineers clone the SDK, fork the examples, watch for the release that supports their model and open issues when the benchmark does not match. That is the evaluation, and most vendors never see it. DemandCompass rolls each one up to the engineer's company, and with a second signal it is the earliest buying window you will see.
ML platform, inference and AI infrastructure roles: the team is building, and the spend follows.
Evidence the account is building agents of its own, from public sources.
For public companies, the annual filing says where AI spend is going. It can lift an account to the line, never across it alone.
A new model server, vector store or GPU provider opens a window next to it.
The install base you can displace, account by account.
The engineer who shipped on your platform before just joined a new account.
Docs, quickstarts and pricing read by people at the same company.
The Decay Rule ages each signal on its own curve. Readiness fades fast, awareness slower. A trial from two quarters ago stops counting.
A new AI infrastructure deployment starts from weights already learned in comparable deployments in its category, where the sample is large enough to trust.
Each week the system checks which signals preceded a real conversation and proposes a change. A person approves it. Nothing changes on its own.
829 signals captured, 150 surfaced to a person, 82% filtered out on purpose.
One deployment, one quarter.