Tools

Things you can judge before you talk to us.

Most of what a consultancy shows you is a claim about work you cannot inspect. Client work is under agreement, and the good parts are usually the parts that cannot be published.

So some of ours is published and running: code you can read and tools you can use, for the same reason we build client work. An argument you can run beats an argument you have to take on trust.

Start here

Live demo

The trustworthiness monitor

The working demo of the trust service, which is middle management for an AI team, shown here on one member of one. A fictional benefits assistant is watched for a month at a time across six stories: a quiet month, a shifting question mix, a stale knowledge base, failing retrieval, a swapped model, and a world that moves underneath the answers. A health card sits on top and every figure carries a stamp saying how current its evidence is.

Live demo

Prevalence audit

The training-corpus measurement method running as a product you can drive. A judge reads a sample, humans calibrate it, and the whole corpus is reported as an interval with a stated error rate.

Open source

Loop Lab

The same agent operating loop, built twice: once by hand, once on a framework, over a shared tool layer. Then both were measured under a real process kill rather than a happy-path demo. If you are deciding whether to adopt an agent framework, this is the comparison nobody publishes.

Also published

A responder subgroup that stops replicating Responder signatures picked from hundreds of candidates on a small discovery cohort, then scored against a confirmatory arm from the same population, a data refresh and the placebo arm, with the rule for when someone gets paged stated before the data arrives. Every scenario result is computed in the page with its interval, so a small cohort shows the wide estimate it is, and each page rule states its false-page rate, measured offline with the same code. Quillmere Therapeutics is invented and every patient is synthetic. An AI line judge that drifts An electronic line judge whose calls shift after a mid-tournament model update, caught from the call stream alone with no ground truth. Synthetic data, and the same question the trustworthiness monitor asks about an assistant. ClearShift: credential-aware shift matching A worked example of credential-aware shift matching for healthcare staffing, with the executive value model behind it. ClinicForce is an invented company and the data is synthetic. Subgroup measurement on clinical text The same measurement discipline on de-identified medical notes, broken out by subgroup so a fairness question gets an answer with error bars instead of an opinion. Brand-safety blocklist scorer Your blocklist is a business decision you cannot see. This scores one against a real news corpus and shows what it quietly kills, so the policy becomes a choice instead of an accident. Contact-center compliance judge A judge that scores call transcripts against a compliance rulebook, then a second pass that audits the judge itself. Fully synthetic data, every number reproducible from a seed. The client-sourcing skill The outbound pipeline this practice runs on, published as a reusable skill under an MIT licence. Value-first, grounded in receipts, and specific about what it refuses to do. We published the machinery rather than describing it. Solar sizing, from first principles Residential solar payback from first principles and public inputs. No vendor arithmetic, no lead capture. Built because every quote in the category is a sales instrument wearing a calculator costume. The Domain-Ramp Workbook A structured way to read an unfamiliar industry in about six weeks, from outsider to contributing in a review meeting. Seeded with a fully worked example from higher-ed enrollment marketing, a domain read for eight years at Sparkroom. Type in your own domain and the workbook starts fresh; nothing typed leaves the page. Metric-SEV console A synthetic demo of a documented metric bound wired to a pager: press play and watch a fairness-parity gap drift past its threshold, fire a SEV-1, and open its own runbook. The dashboard does not change from incident to incident. The on-call rotation does. Courier-elasticity simulator A synthetic simulator built on the supply-elasticity math behind a real Uber marketplace launch. Set a peak-window incentive and watch the lift split into couriers actually bought versus free riders who would have shown up anyway. This runs the math; the client build stayed at Uber. Media-allocation optimizer The GCU media-planning finding, rebuilt on a synthetic fixture: optimizing the upstream metric can make the downstream one worse. A second mode runs the same math against your own budget and cited public benchmarks, and exports a plan. The data here is synthetic; the underlying finding is real. Lantern A safer YouTube for kids, chosen by you. A parent sets what their child can watch in their own words, every video is checked against those settings before it reaches the screen, and a weekly report card shows what was actually watched. Built and run by Jeff, in beta with a small group of friends and family.

Why we run the practice this way

Blue Camel is partly a working test of its own argument. The thesis is that a small shop with modern tooling can own its means of production rather than rent them, and that the interesting question of this decade is who ends up owning the machines.

Jeff wrote that out at length as The Owner's Toolkit for the AI Age, sixteen notes for the person running something small while AI rewrites who needs whom. This practice is one of the receipts for it.

If you are running something small and thinking about the same problem, we would rather hear from you than not. Some of the best conversations we have had this year started that way and were never going to be engagements.

Contact

[email protected]. Peers and fellow travellers welcome, not only buyers.