← All workWeb · 2026

TUMA

A unit manager in municipal elderly care juggles mail, staff schedules, deviation reports and Stratsys follow-up in separate systems. TUMA puts them in one workspace with an AI assistant that prepares the work and asks before anything leaves. I designed and built it as a working sales prototype with a fictional municipality.

Role
Product design · Build
Platforms
Web
Team
Techtify
Made at
Techtify Labs · in-house product
Disciplines
Product design · UI/UX · Accessibility · Front-end
TUMA, desktop
TUMA, mobile

Context

An enhetschef in elderly care starts the day in Outlook, then the scheduling system, the deviation system and Stratsys, and still has to write the APT agenda. Nothing connects, so the manager does the joining-up in their head: a deadline in three days, two urgent mails, one schedule gap.

Who it's for

Municipal staff, starting with unit managers in elderly care. The prototype is shown to municipalities as a sales tool.

Constraints

  • 01It's a prototype: all data is fictional ("Exempelby kommun"), no real systems are connected, and sign-in uses demo accounts until SITHS eID exists. The product says so plainly.
  • 02Public-sector software must meet WCAG 2.1 AA under the Swedish accessibility law (DOS-lagen).
  • 03Work data includes personal data about staff and care recipients, so what the AI sees matters as much as what it says.
  • 04Each source system has to be swappable for the real integration without touching the interface.

My role

I designed the workspace and built the prototype: the five modes, the assistant and its tools, the data layer behind it, and the plain-language information pages about AI, privacy and accessibility.

  • Persona, modes and information architecture
  • Assistant interaction: tools, steps and confirmations
  • Accessible components: dialogs, keyboard paths, focus
  • Planning view, year wheel and daily brief
  • Information pages on AI, BOB, personal data and accessibility

Key decisions

The calls that shaped it

01

The assistant works in the open and asks before sending

Problem
An assistant that can edit a reply and send mail for a manager in elderly care can do real harm if it acts out of sight.
Decision
The assistant works on the draft that's on screen, so you watch every change and can undo it with Ctrl+Z. Each tool call shows as a step in the chat. Anything that leaves the building, like sending, stops at an inline "Godkänn / Avbryt" card. The assistant can't change other parts of the interface directly; each view registers what it will let the assistant do.
Why
Staff stay responsible for what goes out in their name. Seeing the work happen builds trust faster than a promise that the AI is careful.
02

One conversation per mode

Problem
A single chat thread mixes everything: triaging mail, then writing an agenda, then planning the week. The model loses track, and so does the user.
Decision
Dashboard, Work, Mail, Planning and Create each have their own assistant conversation and history. Entering a mode starts a fresh conversation with that page's content as context, and the previous one is kept in that mode's history.
Why
Context matches the task in front of you. Each mode also has its own URL, so links, the back button and bookmarks work.
03

Every drag has a keyboard path

Problem
Planning your day by dragging tasks onto a calendar is quick with a mouse and impossible with a keyboard or screen reader.
Decision
Every task card has a "Planera in" button that opens a small scheduling form, and the day view says "drag tasks here or use Planera in". Dialogs use the native dialog element, and the confirmation dialog puts focus on Cancel first. Status is never shown by colour alone, and motion respects reduced-motion settings.
Why
DOS-lagen requires it, and municipal staff include people who rely on keyboards and screen readers. The keyboard path is also faster for anyone planning many tasks.
04

Explain what the AI sees, in plain Swedish

Problem
Before a municipality can use an AI assistant with care data, the people using it need to know where their questions go. That is usually buried in an agreement nobody reads.
Decision
Information pages inside the product explain the AI, personal data and accessibility. One page walks through BOB in five steps: names and personal identity numbers are swapped for placeholders before a question leaves the organisation, and put back in the answer. The page also says that the demo doesn't use BOB, because all its data is made up.
Why
Being honest about the prototype's limits makes the sales conversation more credible, not less.

Responsive

Five modes, one assistant

Each mode has its own AI conversation on the right and works from 360 px up. All people and messages are fictional demo data for Exempelby kommun.

TUMA: Overview, desktop

"Your day in brief": prioritised mail, today's tasks and the week, in that order. · 1440 px

Design system

The system

A light, warm glass workspace with one sky blue for action. Actions use a darker shade so text keeps AA contrast.

Sky#0284C7
Sky 700, actions#0369A1
Gradient light#EFEFEF
Gradient warm#D7DAC6
Text#111827
Night#0F172A
Aa

Poppins

Interface

Aa

Figtree

Headings on information pages and the assistant

  • Amber marks anything waiting for your approval
  • Light and dark themes, saved per user
  • Layouts work from 360 px wide; multi-column views stack on small screens
  • The interface is in Swedish

Outcome

Where it landed

TUMA is a working prototype on fictional data. It runs on Vercel behind demo accounts, one per audience, so access can be revoked per meeting. Real integrations (Microsoft 365, scheduling, Stratsys), SITHS sign-in, role-based access and audit logging of AI use are designed and documented but not built. Hosting for production is still open, since municipalities often require Sweden or the EU.

5modes in one workspace
AAWCAG 2.1 target under DOS-lagen
Next caseChargeNetAn EV charging app: find a charger, start it, and plan the stops on longer trips.