Saturday, 7 November 20265 stories in the last 30 days

The Journal

What LedgerTM shipped, and how we build itBy Jeremy Rich
Templates

A test bank where the customers actually do things

A storyboard is a cast of made-up customers, each with a life written out day by day, that plays through the real API every time the test bank closes a day. Sixteen cardholders now live in the UK card issuer template, and one of them reports a lost card every month.

Read the article →
How we build

Measuring the factory, not just the product

We log what the development process itself does, every build, gate, test, question and token, in plain files on the machines that do the work. It showed us a commit that rebuilt 121 projects for no reason, and that of sixty failed gates in one week only seven had found a real defect.

Read →
How we build

AI writes tests faster than we can run them

Our integration tests doubled in fifteen days and the merge gate went to half an hour a change. We now build once, ship the build as an image, and run the tests in parallel on rented Ubicloud runners. It's better, and it still isn't where we want it.

Read →
How we build

The test is written before the code, and by somebody else

Every design that changes how the product behaves now carries its specification as plain Given, When, Then scenarios. A separate AI author turns them into integration tests before any code exists, and Bobcat runs them and checks that the test still says what the scenario says.

Read →
How we build

What the Linux box does while I'm asleep

At one in the morning the box runs every test we have, measures coverage, and walks every template's sample book through three months on the last release and on today's code to see whether they still agree. Then a night shift fixes the small things, and at seven there is a report. For one week in November it fixed nothing at all.

Read →
ProductNext release

Let your own agent read the ledger

The MCP door lets your organisation's own AI agent read LedgerTM over the Model Context Protocol, with a person's permissions and a record of every call. Anthropic's GL Reconciler wants a read-only subledger to reconcile against, and this is one.

Read →
Product

Point Power BI at the book

Reporting tables give Power BI, Tableau, Excel or any SQL tool a live, read-only view of the book, shaped for reporting. Each account type gets its own table, built from its configuration, so a new product can be reported on the moment it's published.

Read →
Product

Context-sensitive help, on steroids

Ask LedgerTM how to do something, or ask about the account in front of you. The assistant knows which screen you're on, reads only what you may read, and links every figure to the record it came from.

Read →
How we build

A third faster, and fewer false alarms, in our integration tests

We trialled Bobcat, a new open-source test supervisor from the makers of Marten and Wolverine, on our slowest suite. It cut the run by 36% and gives us an honest answer to the question every red test raises — is this a real failure, or a flaky test?

Read →
Product

Did the client get the letter? One place to answer it

Statements, notices and alerts now go out through one correspondence service, by email, SMS or post, with WhatsApp next. Each client chooses how they hear from the bank, and every send has a recorded outcome.

Read →
TemplatesIn progress

Corporate banking that charges for every payment

A corporate client signs a service agreement, is enrolled in services, and is charged per payment at an agreed price. Metered usage closes each month into a charge file the bank's own billing system reads.

Read →
Templates

A microlender, from a template

Short-term unsecured loans to people, collected by debit order, held to the National Credit Act's fee ceilings, with arrears, promises to pay and provisioning built in.

Read →
How we build

Five lanes per developer, and a coordinator between them

Each developer runs several lanes of implementation work at once, each an AI-assisted session in its own workspace. A coordination session watches the space between them, routes their questions to one place, keeps the status honest and loses nothing. It never writes code.

Read →
How we build

Grilling a design before we build it

Before a design is built, a workflow of AI reviewers reads it from four seats, the clerk, the auditor, the financial director and the salesperson, checks every question against the code, and hands us a ranked agenda of the questions that would change the design.

Read →
How we build

We audited our own codebase, and published the grade

A five-stage, principal-level audit of the whole engine, run by the same AI-assisted loop that builds it, graded A minus, with every serious finding fixed in seven sessions and the ones we chose not to fix written down.

Read →

Written by Jeremy Rich. Each story is stamped with the time its work landed, South African time.