Work Case study My own product

LumenPort.

A local-first browser testing workbench. A controlled browser, driven from a paired web page, with a chat assistant answered by a model on my own hardware, so the pages I test never go to a cloud AI.

tested independently checked

01 What it is

The workbench

Browser
Chromium or Firefox, through Playwright
Control
A paired web page; desktop approval pop-ups
Permissions
Per-site profiles: read, click, fill, submit
Assistant
A local 120-billion-parameter model on a home AI server

LumenPort puts a real browser under explicit control. Each site gets a permission profile that says what the workbench may do there: read only, click, fill in fields, or submit. Sites can be saved. Actions that cross a line raise a desktop approval pop-up. For authenticated crawls, the operator signs in with their own account inside the controlled browser, rather than the tool being handed a login of its own.

From there it crawls a site into maps by domain, records and replays scripts, and compares pages. The chat assistant that helps read and reason about what it sees is answered by a local 120-billion-parameter model running on a home AI server. Nothing a client shows me is sent to a cloud model to be understood.

02 Why it exists

Permission is a mechanism, not a promise.

Two things go wrong with AI-assisted testing tools. They act on a site in ways nobody authorised, and they send what they see to someone else’s servers. LumenPort treats both as design problems rather than policy problems.

  • Per-site permission profiles mean a read-only site cannot be clicked, let alone submitted, no matter what the assistant suggests.
  • Approval pop-ups on the desktop put a person at the boundary where a suggestion would become an action.
  • A local model means the privacy rule is enforced by where the computation happens, not by a terms-of-service page.

It is the working example of the local-first AI setup I offer: the same rules, applied to my own tools first. With a client’s permission I also use it on client work, locally: it maps the product before any Playwright checks are written, and the client’s data stays local.

03 Engineering practice

Every change is reviewed by a model that did not write it.

Every change to LumenPort gets an independent review by a different AI model, and fixes go round until the reviewer says ship. independently checked

One example: the tripwire that cried wolf

A safety tripwire, there to stop the workbench from doing something it should not, started firing on something harmless. The symptom-level fix would have been to loosen the threshold. Instead the false positive was traced to its cause, reproduced deliberately, fixed at the mechanism, and pinned with 28 regression tests so it cannot return unnoticed. tested

That is fix mechanisms, not symptoms, applied to my own tools. It is also why the tripwire can be trusted to stay strict.

04 What this means for your project

What you can take from it.

Your pages stay on the machine that tests them

When I explore your product, the understanding happens locally. If a cloud model is ever useful, you will be asked first, and the material will be anonymised.

Authenticated testing under the operator’s own account

For pages behind a login, the operator signs in with their own account in the controlled browser, so testing does not have to begin with handing a tool a password. How access to your product is arranged, and what is kept where, is written into the charter before anything starts.

Independent review as a service

The same second-model review loop can be set up for your team’s changes. It is part of the automation and AI setup service.

Next

The other case, or a conversation.