A back office for review sites where the evidence and the rollback both matter
An operator application that takes a candidate domain, gathers bounded evidence, extracts and scores what it found, passes it to review, and then renders a static release that can be activated or rolled back.
Review sites fail in the same two places over and over: nothing keeps the evidence behind a claim, and nothing lets you undo the release that published it. I built one control plane for both.
How it works
A candidate domain comes in, bounded evidence is gathered, and each kind of evidence is extracted into a typed value and scored instead of being summarised into a paragraph. The result goes to a reviewer who can accept, hold or reject it, and the revision loop runs until a decision exists. Only then does the system render a static release, validate it and activate it — keeping the previous release, so a bad activation can be rolled back.
What I built
- An Elixir, Phoenix and LiveView office on PostgreSQL, with Oban running the background work. All seventeen worker types were implemented at the recorded point.
- Go and Rod sidecars for crawling and screenshots, and bounded Lighthouse sidecars, all behind shared target-safety controls so a worker cannot be pointed at an arbitrary host.
- A Rust renderer that turns a frozen manifest into a validated static release, with retention rules and an encrypted write-only store for connector credentials.
- An explicit release boundary: the workers never publish, and activation is a human decision with a recorded actor.
Where it stands
The tooling is private, and the production rollout and the human activation were still gated when I recorded this. Nothing here claims an unattended pipeline or a commercial result. What the work is worth is the shape of the flow and the evidence it keeps, and that is what this page describes.
Have a system like this?
Describe the workflow, the data or the system you need built. Scope and any quote are handled privately after the enquiry.
Published