Themes, Variants, Formats
A three-layer brand system. Structure, skin, and generation recipes are managed centrally and flow to every user instantly — consistency by default rather than by discipline.
Case study 01 · for Planfocus Consulting
AdviceStudio
Turns a dense 40-page compliance document into a client-ready interactive presentation in minutes — without the client’s personal data ever reaching the AI.
The problem
The advice was good. The meeting wasn’t.
Financial advisers must produce a Statement of Advice — a long, legally compliant Word document setting out their recommendations. It is correct, it is required, and it is close to unreadable. It then gets presented in what should be the most valuable hour of the client relationship.
The alternatives were building a presentation by hand every time — hours per client, inconsistent and off-brand — or paying a paraplanner or designer to do it. Neither scales across a practice, let alone a licensee with hundreds of advisers.
What we built
Upload a document, get a presentation
01
Upload
The adviser uploads the .docx they have already written. No new inputs, no re-typing, no separate content step.
02
The AI builds it
The document is read, structured into sections, and laid out in the organisation’s house style. Charts are generated from the actual figures in the advice. The structure adapts to what the document contains rather than forcing everything into one template.
03
Review, edit, present
A full inline editor — text, tables, images, section order, undo and redo. Then presenter mode with per-slide notes, or a clean PDF export to send afterwards.
The hard part
Personal data never reaches the AI.
Before a single word is sent, a dedicated redaction engine strips names, dates of birth, addresses, phone numbers, emails and account numbers, replacing them with neutral placeholders — including a purpose-built Australian address recogniser layered on top of the general model. The AI only ever sees a de-identified document.
The real values are stitched back in inside the application, afterwards. And before sending, the de-identified text is re-scanned: if any identifier slipped through, the system stops and refuses to send rather than risk a leak. Fail-safe, not fail-open.
That pattern — de-identify, process, reinstate, verify — is how we approach any AI system handling data that can’t leave the building.
Also built
What makes it work across a whole organisation
Presenter mode
Per-slide notes: an opening line, the number to land on, an anticipated question with a prepared answer, and a transition. Visible only to the adviser.
Review workflow
Draft, in review, presented, archived. Reviewers leave comments and visual mark-up before anything reaches a client.
Security controls
Tenant isolation, two-factor authentication on by default, single-session enforcement, encryption at rest with managed keys.
Audit trail
An immutable events log of user actions that the organisation can rely on for governance.
Data residency
Hosted in AWS Sydney, so client data stays in the jurisdiction the regulator expects.
Stack
What it runs on
- ApplicationPython 3.12, FastAPI
- DataPostgreSQL 17 on RDS, Alembic migrations
- ComputeECS Fargate, two services behind an ALB
- ExportHeadless Chromium via Playwright for PDF rendering
- BillingStripe, with per-organisation plans and seat management
- DeliveryCodeBuild to ECR, rolling ECS deployments
- Regionap-southeast-2 (Sydney)
Worth being precise about
What AdviceStudio is not
It does not write the advice. The licensed adviser authors the document; AdviceStudio transforms a finished, compliant document into a presentation. That distinction matters enormously in the compliance conversation, and it is a strength rather than a hedge — the human professional remains the author of the advice.
It is also not a generic slide tool. It is purpose-built for a specific document type and a specific meeting.
Data that can’t reach a model?
That is the problem we are best at. Tell us about yours.