Case study — Public help
Built on caveauAI · a Blue Note Logic founder project, not a client engagement
Purple Askepott: help that shows where it comes from.
A free help site for women and men living with domestic abuse, built on caveauAI in November 2026. It points people to checked help in Norway, Ireland, the UK and the US, and its assistant answers only from real help pages, then shows which ones it used.
Need help now? Purple Askepott's Get help now page lists checked helplines by country.
Why it exists
Help exists. Finding it safely is the hard part.
Someone living with abuse often looks for help in a hurry, late at night, on a phone someone else may check. The help is real (helplines, crisis centres, legal aid, health services, the police), but it is spread across many websites, in different languages, and every country arranges it differently.
In that moment a wrong phone number is worse than none, and a site that tracks its visitors can put them at risk. Purple Askepott was built to be the opposite: one calm place that says where to go, says it plainly and leaves no trail.
For caveauAI it is the hardest kind of deployment. The readers may be in danger, the facts have to be right, and the privacy promise has to hold in production, not only on a slide.
What a visitor gets
One site, four doors.
The site speaks to women and men on equal terms, with its own pages for children and for the people around them. Help is also sorted by topic for each country, and a Quick exit button on every page jumps straight to a weather site.
For children
Written for children and young people, with help they can reach themselves.
Open the page
Plates: AI-generated illustrations (Codex image generation), from the pages they open.
What it proves about caveauAI
Five properties, checked in production.
- Safety before any model callA rule-based check reads every question first. If it points to danger, the visitor gets checked local help, and no AI answer is generated at all.
- Answers that show their sourcesEvery answer lists the pages it was built from, with links, so the reader can check it rather than trust it.
- Country-scoped retrievalThe visitor chooses a country. Search is limited to that country's sources before anything is ranked, so services from different countries never mix.
- No question text stored or loggedQuestions are sent without the visitor's IP address and are kept out of logs and caches. Misuse is limited with a one-way code that is deleted after an hour.
- Facts that are checked, then re-checkedEvery phone number comes from one verified list that the build enforces. Every service was checked against its official page, and a monthly check flags any page that changes.
A visitor asks a question and chooses a country. A safety check runs first: if the question shows signs of danger, the visitor gets checked local help and no AI answer. Otherwise only that country's help pages and official guidance are searched, a short answer is written, unchecked phone numbers are removed, and the answer is shown with the pages it used and matching services.
Diagram source
flowchart LR
Q["A visitor asks a question<br/>and chooses a country"] --> S{"Safety check:<br/>signs of danger?"}
S -- "yes" --> H["Checked local help first<br/>no AI answer"]
S -- "no" --> R["Search only that country's<br/>help pages and official guidance"]
R --> A["A short answer<br/>in plain language"]
A --> C["Unchecked phone numbers<br/>removed"]
C --> V["The answer, the pages it used,<br/>and matching services"]What we measured
Before and after the assistant's search was rebuilt.
Built and tested, not yet switched on
What comes next is already built.
Each of these has passed its tests and stays off until it is ready to open. How each one works.
Members
An email address and an 18+ tick, nothing more. Sign in with a short-lived code; download everything or delete the account at once.
My notebook
A private notebook encrypted in the visitor's own browser with a passphrase only they know. The server keeps text it cannot read.
The weekly update
Off until chosen, written by people from what is already on the site, with a one-click stop in every email.
Creator Studio
Publishing to Purple Askepott's own channels, where every post needs an approval code and a global switch can stop all posting.
Comment insights
Public comments, stripped of names, links and contact details before anything is kept, and deleted after 28 days.
How it was built
AI agents did the work. A person approved every release.
Dave Gilligan built Purple Askepott with AI coding agents, Claude Code from Anthropic and Codex from OpenAI, each working on its own branch. Nothing reached production without passing its tests and a person's approval.
The site went live on 4 November 2026. Help by topic for four countries and an Irish-language beta followed within days, and each change to the assistant's search was measured before it shipped.
AI coding agents work on a branch. Tests and a deploy gate run, a person reviews and approves, and the release ships with a snapshot. Live questions and the test set are measured, which feeds the next change. If anything is wrong, the release is rolled back to the snapshot.
Diagram source
flowchart LR
B["AI coding agents<br/>work on a branch"] --> T["Tests and<br/>a deploy gate"]
T --> P["A person reviews<br/>and approves"]
P --> D["Release with<br/>a snapshot"]
D --> M["Measure: live questions<br/>and the test set"]
M --> B
D -. "if anything is wrong" .-> R["Roll back<br/>to the snapshot"]Limitations
What this case study does not claim.
- Purple Askepott is not an emergency service and gives no legal or medical advice. It points to the people who do.
- The assistant can be wrong. That is why every answer shows its sources.
- The Irish version is a beta.
- The figures above are our own estimates from our own test set.
- This is a founder project. It shows what caveauAI can do; it is not a client result.
Where next