Axiom Orbit. Findings you can stand behind.

Software we engineered to find the work hiding in a client’s website, and to prove each finding before anyone sees it.

Case studyRead on Or experience it as a 3D story

Axiom Orbit

Verified before it is shown. Axiom Orbit is the product we built to find the billable work hiding in a client’s website. It is invite-only for now, and a working example of how we put AI to work without trusting it blindly.

At a glance

Kind: Proprietary software product. Not open to the public.
Built by: Axiom International, led by Aidan Magee
Built for: Small web and digital agencies
Status: Deployed to production behind invite-only sign-in, with no billing yet
Built with: TypeScript, React Router and Cloudflare Workers with D1
Tested by: More than 900 automated tests, and a 24-site corpus replayed offline (September 2026)

The challenge

Every audit found something. Almost none of it was work you could sell.

An agency wants to know what it could sell an existing client. Audit tools answer a different question, and when they do answer, they are sometimes wrong.

Noisy

On a 24-site test set, routine site-health findings outnumbered sellable ones four to one. A broken link led the list while a missing service page sat six rows down.

Unchecked

A tool that says a page is missing has said nothing until the site is checked. A wrong finding costs an agency its credibility with the client.

Costly

Asking a language model to judge every candidate is slow and expensive, and its own confidence scores did not separate a real service from a marketing claim.

The solution

One private product that finds the work and proves it.

Axiom Orbit crawls a client’s website, applies deterministic rules, verifies each finding against the crawl, asks the model for judgment only on what survives, and turns what is left into a priced proposal an agency can share.

Crawler

Reads a client’s site within a strict budget, identifies itself, and obeys robots.txt. A page it is not allowed to read is recorded as blocked, never as clean.

Rules

Eleven deterministic rules. Three commercial rules, such as a missing service page or a broken conversion path, lead the queue. Eight site-health checks sit below them.

Judgment gate

A classification, not a vibe: is this a distinct, sellable service? Anything the model is unsure about, or fails to answer, is rejected.

Monitoring

Opt-in per client and off by default. Scheduled checks run in bounded batches, and a weekly digest reports what changed in billable terms to agencies that have it enabled.

Proposals

A finding becomes a priced proposal from the agency’s own catalogue, shareable through an expiring, revocable link.

Workspaces

Each agency has its own workspace. Its data, limits and history are separate from every other agency’s.

Engineering

Where the engineering is.

A selection of what we built and chose carefully. Each rule below is enforced in the code, not left to convention.

Absence is verified, not asserted.

Before Orbit tells an agency that a client has no page for a service, it checks the crawl for it, and it makes no claim at all from a thin crawl. We found this the hard way: auditing our own engine against real sites showed a third of its commercial findings were false, because a page linked from every page of a site had been scored below a threshold. We fixed the cause and pinned it with a regression test.

AlsoEvery one of the 30 findings on the test corpus was checked against the cached pages by hand.

The expensive step comes last.

Findings pass deterministic rules, an evidence threshold, a billability check and suppression of anything already covered, sold or dismissed. Only then is the model asked, and only for a classification. Anything but a clear yes is rejected.

AlsoA platform-wide daily ceiling, and a deploy-time check that prints the worst-case paid day, keep the bill bounded.

Honest about what the model knows.

A measured calibration run showed the model’s own confidence numbers did not separate a real service from a trust claim, so the field was removed rather than left looking meaningful. The confidence on a finding is the evidence strength the rules computed.

AlsoA 37-case adversarial benchmark covers true services, trust signals, promotions, generic claims and ambiguous edges.

Tenants cannot see each other.

Every read and write is scoped to a workspace at the repository layer, and the database refuses cross-workspace writes through composite foreign keys. Isolation is tested at the repository, route and unauthenticated level.

AlsoThe base URL for sign-in comes from trusted configuration, never from a request header, and sign-up is invite-only by default.

A good citizen on the web.

The crawler identifies itself and obeys robots.txt, including when it verifies that a page is missing or a link is broken, because ignoring it there would manufacture false findings.

AlsoMonitoring is opt-in per client, runs in bounded batches with one owner per client, and backs off after failures.

How it is built

Interface: React Router v7 on Cloudflare Workers
Storage: Cloudflare D1, with versioned migrations
Sign-in: Better Auth with email verification and database-backed sessions
Engine: A portable domain engine in pure TypeScript with no framework dependency, testable offline at no cost
Integrations: A language model for the judgment step, and transactional email

Where it stops

Not open yet

Orbit is deployed behind invite-only sign-in. It has no billing and has not been opened to outside agencies; a live sign-in and email check is a required gate before any pilot.

Not every site can be read

On the 24-site test corpus, 19 sites were analyzable. The rest were reported as unreadable rather than as clean.

Findings are evidence, not forecasts

A verified finding shows what is on the client’s site, not that the client will buy. Orbit prices work from the agency’s own catalogue and makes no revenue prediction.

For your business

The same engineering, pointed at your problem.

Axiom Orbit is ours, so every decision in it was ours to make. These are the problems the same thinking applies to.

Your tools produce volume, not decisions.

We build the layer that decides: deterministic rules first, then evidence, then a narrow judgment, so what reaches your team is worth their time.

You want AI in the workflow, and it is expensive and overconfident.

We put the model last, ask it narrow questions, fail closed when it is unsure, and cap the spend in code.

Several clients share one system.

Separate workspaces enforced by the database itself, not by hope, with isolation tested at every layer.

Off-the-shelf software almost fits.

Where a standard product does the job, we will say so. Where your own process is the advantage, we build software around it.

Custom software

Tell us how your business runs, and where the software you have stops.

contact@axiominternational.ca

Kitchener–Waterloo, Ontario, Canada

All work

All contact details