The problem

Most revenue teams already own the data that would tell them which accounts to work. It sits across the CRM, the marketing automation platform, the product telemetry, and the intent vendor nobody has audited in two years. What they lack is a scoring layer that turns it into a ranked list a rep will actually open. The usual answer is a points-based model built in the CRM: fifteen points for a demo request, ten for a pricing page view, decay after thirty days. It is explainable, which is why it survives. It is also mostly wrong, because the weights were set by consensus in a meeting rather than derived from what closed. The second answer, buying a scoring vendor, replaces consensus weights with a model trained on somebody else's outcomes. It scores your accounts against the average of a category you may not sit in.

What ax1om does

Per-customer models, not a shared one. Each customer gets a LightGBM model fit on their own conversion history. The floor is roughly fifty closed opportunities. Below that there is nothing to learn from, and saying so is more useful than shipping a confident-looking score. Explanation attached to every score. SHAP values put the why next to the number. This is not a nicety: a rep who cannot interrogate a score will route around it, and the first time someone sees a top-tier account justified by a careers-page visit, you have learned something no validation metric would have surfaced. Timing, not just fit. In-market timing inference separates the account that looks good on paper from the account that looks good this week. Fit tells you who to sell to; timing tells you when, and the second question is the one a rep actually has. Delivered where the work happens. CRM writeback puts scores on the record, plus a metered scoring API for teams that want to consume them elsewhere. A dashboard nobody opens is not a product.

Architecture

FastAPI on Cloud Run behind a Next.js front end, Stripe for metering and billing, LightGBM with SHAP for the model layer. Past 200 production deploys. The GTM side runs on infrastructure built rather than bought: an event router that replaced Clay and n8n, PLG funnel instrumentation, and in-house customer success and email modules. Total stack cost is under $100 a month. That number is a design constraint, not a brag. A scoring platform whose unit economics only work at enterprise pricing cannot serve the fifty-to-five-hundred employee companies that need scoring most, because they are the ones without a data science function. Keeping the floor cheap is what keeps the market open.

Status

In enterprise beta.

Why it exists

The GTM engineering conversation is full of people describing systems nobody has built. ax1om is the counterargument: a working system, built solo, with the design decisions and the dead ends both on the record.