All work
2026
Why we built the Vaultia Knowledge Base
Why AI tools fall short with regulated documents, and what Vaultia’s Knowledge Base does instead.
Vaultia
Product design
In development
Client
Vaultia
My role
Product, design and build
Timeline
9 weeks
Year
2026
Vaultia started with a simple problem. Healthcare teams wanted to use AI with their own documents: approved materials, brand guidelines, research. But public AI tools ask you to put those documents in someone else’s cloud. And the tools built for documents treat them as a pile of files to search. For regulated work, neither is good enough. The Knowledge Base is what we designed instead. It’s part of Vaultia, the private AI platform I co-founded.
The challenge
A pile of files doesn’t know what it’s holding. In a typical AI project or shared drive, the AI searches your documents for matching words. It can’t tell an approved leaflet from last year’s draft, a patient piece from one written for doctors, or one brand from another, so it answers from the closest match.
Some documents can’t be left to a search. An approved terms list or a rules checklist isn’t something to search through. It’s the thing you check against. If a search for an approved term finds nothing, the AI can’t tell “not approved” from “not found”, and both read as “no problem”.
Not everyone should see everything. Most tools give access to a whole project, or none of it. An agency working on competing brands, or a company with confidential work, needs walls inside one library.
Documents go out of date. Guidelines get replaced and claims get withdrawn, and a typical AI tool keeps using the old version without saying so.
The approach
Your documents stay yours. Vaultia is built to run inside each client’s own AWS account. The documents are stored there, in Australia, and when they’re used in Vaultia, the AI that reads them runs in Australia too. Nothing is used to train a model, and the company that makes the model can’t see it.
Every document carries its own record. What it is, which brand and audience it’s for, whether it’s approved, who owns it and when it’s due for review. Each company sets up the fields that matter to it. When a document is added, AI reads it and drafts the record, quoting where each detail came from, and a person approves it. That record is what lets the AI be told to search only approved, current material for one brand and one audience.
Rules are read in full, never searched. They sit apart from everything else, in storage the search can’t reach, and the AI reads them in full every time.
Walls inside one library. The Knowledge Base can wall documents by brand, client or team. The AI only answers from what the person asking is allowed to see, and everything else doesn’t exist for them. A new person starts with the least access, and if the access check fails, the answer is no.
Nothing quietly goes stale. Every document has an owner and a review date, and every version is kept, with only one live. When a rule falls overdue, the AI flags it in every answer, or stops using it until it’s reviewed, whichever the company chooses. The history shows who changed what and when, so “what did the rules say in April?” has an answer.
It works with the AI tools your team already uses. The Knowledge Base is a service in its own right, and agents like Claude and ChatGPT can connect to it. They get answers, not the library: only the passages a question needs, and only what that person is allowed to see. The originals stay in your account. What’s sent back is then handled under that tool’s own terms, so for the most sensitive material, use Vaultia’s own chat, or an agent that runs in your own AWS account.
How I built it. I ran the build the way I’d run a development team. I made the product decisions and the design calls and wrote the specification, and Claude Code wrote most of the code against it. The specification came first: a written contract for everything the backend has to provide. More than 1,800 tests and 12 automated checks run before every push, and 15 decision records explain why the main rules exist. Nine times the process caught its own mistakes, and each fix became a new rule or check.
The outcome
Where it stands. The front end and the specification are complete. Everything above is designed, built and tested in a finished front end, running against a reference backend that applies the same rules, with invented content and no client data. Connecting it to live storage and AI search inside the client’s AWS account is the next phase, with my co-founder.
Tests guarding the product’s rules
1,800+
Decision records explaining why
15
Mistakes the process caught and turned into checks
9
Next project
A report that counts what the ads actually earned
Based in the Blue Mountains
Open to contract work
If you’re working out where AI fits in your team, I’d like to hear about it.
Email or LinkedIn is the easiest way to reach me.