In Azure, in South Africa
The site is an Azure Container App in the South Africa North region. It sleeps when nobody visits and wakes on the next visit.
For a business owner deciding whether to trust us with a website and its agent. Every line says what the code does today. Where something isn't built, it says Not yet.
One Azure environment in South Africa. The site is the only thing in it with a public address.
The site is an Azure Container App in the South Africa North region. It sleeps when nobody visits and wakes on the next visit.
The internet reaches the site on port 443 with a managed certificate, and plain http is refused. Nothing else we run in Azure has a public address.
Every visit is https. The site's logs sit in Azure Storage, which Microsoft encrypts at rest with 256-bit AES, and that can't be switched off.
Every key lives in Azure Key Vault. The site reads it by reference with its own managed identity, and a release signs in to Azure from GitHub without a stored password.
The site runs in its own container from its own image. It runs no shell and builds nothing. It reads its docs and passes on a visitor's details.
What the site may call out to is Azure's default. No rule of ours narrows it yet.
The site's image doesn't name a user of its own yet, so it runs as the image's default user.
Under POPIA, South Africa's privacy law. The privacy page has the detail, and the table names everyone who handles data for us.
The site holds a chat in memory while it runs and forgets it when it restarts. For each answer it writes down the cost, the page and the tools used, never your words.
The agent shows what it will send and sends it once, when you press Send it. Nothing reaches us before.
You may ask what we hold about you, have it corrected or deleted, and complain to the Information Regulator.
The table below names each company that handles data for this site, what for, and where.
The logs and the record of costs stay on the site's file share. There is no deletion schedule yet.
| What for | What they see | Where | |
|---|---|---|---|
| Microsoft Azure | Runs the site, keeps its keys and watches it | What the site runs, and its traces without your words | South Africa North |
| Anthropic | Its model writes the answers | The messages of a chat | United States |
| Resend | Sends us your details by email | Your name and contact, after you press Send it | Outside South Africa |
| Cloudflare | Points the domain at the site | Only where a name points. Visits don't pass through it | Worldwide |
| GitHub | Holds the code and releases it | No visitor data | United States |
The rules that matter are code, tested without a model. The prompt says what we intend, and the code makes it so.
An agent's tools are fixed when a chat opens. The site's agent gets its docs, the site, and a way to pass on your details, and a visitor can't ask for more.
Your details go out only on your yes. A skill an agent writes for later waits until a person approves it.
Nothing the model writes reaches a shell. A fetch refuses private addresses and obeys robots.txt, and secret files are never read.
At most 10 rounds of tool calls for each message. A result that is too big is summarised, and a call cut off midway never runs.
Each request to the model is priced in a ledger, and each step of a chat is traced in Azure's monitoring. One address opens at most 20 chats a day.
Agents are checked on evals written in C#, scored by code and by a model as judge, with a person reviewing what the judge flags.
A suite of injection attacks, checks on what the agent says, and a log of every action are planned, not built.
Every merge runs the unit and integration tests. Live evals cost money and are run by hand, so they don't hold back a release yet.
How a change reaches the site, how it would be rebuilt, and what is yours to take.
A merge runs the tests, then releases the site as a new revision from an image with its own tag. Earlier images stay in the registry.
One script releases this site and the framework's own from code. The environment they run in was built once from code that is being retired, so rebuilding it from nothing isn't one command yet.
A site is its own project. Its words, brand, pages and data sit apart from the framework, in code that can be handed over.
The file share that holds the logs and the site's saved content has no backup set up. Chats aren't stored, so no visitor's words can be lost.
The site runs in one region, and a sleeping site takes a moment to wake on the first visit. There is no uptime figure yet.
The pieces to rebuild everything exist. A written, timed walk-through of a full recovery doesn't yet.
A reviewer's list of what a business should see before it pays, and where each item stands today.
A written agreement for when we process your customers' data on your behalf.
Every account already carries its business, but no test yet proves one business can't reach another's data.
Every visit is https, and Azure encrypts what the site stores. Where it runs, above, has the detail.
Not set up yet. If something breaks, above, says why.
Chats aren't kept. The logs and the record of costs have no written period yet.
Anthropic's model writes the answers. Every company that handles data is named under your data, above.
A written plan for what we do, and whom we tell, when something goes wrong. POPIA asks that the Information Regulator and the people affected are told.
Customer terms, a service level and who answers for what are not written yet.
No outside tester has attacked the site yet.