Home › For leaders
For leaders · business and technology
What's going on here, in one minute
| The problem | Companies are putting AI agents in front of customers and AI on top of their data. Most of what goes wrong isn't the model. It is everything around it: the wrong context, a retry that charges a customer twice, a file that arrives twice and doubles revenue. |
|---|---|
| What we do | Build, in the open, the checklists, test worlds and working code that show whether an AI system or a data platform is ready, and teach engineers to build them. Free, with no product to sell. |
| What exists today | Two open checklists (118 things an AI application must have, 118 it must prove). A test world that judges an agent by what it changed, not what it said. A full reference agent that passes all four quality gates. A claims assistant deployed on Azure. A 167-page course and ten simulators. |
| Proof so far | 180 planted attacks, all held. On 2,472 recorded runs from an outside benchmark (τ²-bench), our verdicts matched theirs on every rule-breaking change in two of three domains. 19 data problems recorded on real Spark and Delta. |
| What isn't done | Tests use a scripted model, not live models run many times. The test worlds haven't been checked against real systems yet. The data-platform test tool (DataTwin) is designed, due early 2027. |
| What it costs you | Nothing. Code under Apache 2.0, specifications under CC BY 4.0, no sign-up. Your teams can use all of it as it is. |
Questions leaders ask
Each answer is two lines, with the evidence one click away.
Is our AI agent ready for real customers?
Ready means it keeps every promise it owes, in every situation you can name, under the faults you know about. Ask your team for that list and the evidence for each item, not for a demo.
Should we build our agent or buy one?
A one-page guide: what you must own, what you can rent, and the questions that decide it.
What will it cost to run?
One conversation with our reference agent costs under a tenth of a cent. Agents that cost $4 a question usually make the same nine mistakes.
How do we know it's paying off?
Count the work it removes and the errors it adds, before you scale it.
What can go wrong, and what stops it?
57 common guardrails, each marked must or should, where it sits, and the test that proves it holds.
Is the model the risky part?
Mostly not. In our reference agent the code that calls the model is about 3% (817 of 24,631 lines). Failures live in the other 97%, which ordinary engineering can test.
Can we trust the numbers on our dashboards?
Only if someone checks them against the truth, not against last week. A file that arrives twice doubles revenue while every job stays green.
Why is our data platform bill so high?
The same ten leaks show up again and again, from clusters that never scale down to jobs that rewrite everything.
Which vendor or platform should we choose?
Decide the design first, without product names; then map it to your cloud. Our specifications name no product, and each cloud gets its own binding.
How do we get our team ready?
Free paths for six roles on every major cloud, from beginner to advanced. Point each person at their role and cloud.
How do we build many agents without starting from zero each time?
A small centre of excellence writes the specs and builds the shared assets once. Every agent is then built, tested, released and run from them, and each one makes them stronger.