The Product Journal by Nova Skye Nova Skye · Technical product manager Contact

Feature · 02 · SettleSMART

  1. Problem
  2. Bet
  3. Evidence
  4. Decisions
  5. Result

Helping newcomers get answers they can trust

A settlement chatbot, a moderated multilingual group chat and a private support chat, piloted inside a research study in Alberta from June to November 2026. Presented at the AAISA Settlement and Integration Conference in October 2025, with a poster at Pathways to Prosperity on 23 November 2026.

SettleSMART on a phone in English: Choose your language, with a note that translations are made by a computer
The same screen in Farsi, with the whole layout flipped right to left
The first screen, in English and in Farsi. Thirty-two interface languages, five of them right to left, each machine draft labelled as such until a native speaker has reviewed it.Fig. 01
Stage
Pilot · June to November 2026, inside a research protocol with ethics approval
My role
Discovery, requirements, roadmap, decision log, development, testing, deployment
Team
A full research team and a software development team
The bet
Reliable information in the newcomer's language at any hour, so staff time goes to the cases that need a person
How we judge it
Return visits, use outside office hours, and answers checked against the organization's own
Technology
Web app · cloud database hosted in Canada · Azure AI for chat, translation and speech
Section The story Marginalia · the PM layer
1

The problem

Settlement workers answer the same questions many times a day, in several languages, during office hours, while newcomers wait their turn to ask things like "where do I register my child for school?" The knowledge exists, spread across more than sixty organizations' websites, program pages and PDFs, but the person asking has little time, a lot at stake, and often a new language.

The goal was a tool that increases capacity without lowering trust: reliable information, in the newcomer's language, at any hour. It had to work inside a research protocol at the centre. Data had to stay in Canada, every AI, translation, voice and email action was a paid call, and many users were new to Canada and often not working in English.

2

The bet

The hypothesis is that reliable answers in the newcomer's own language, at any hour, will let staff time go to the cases that need a person without lowering the trust the organization has earned.

It works if people come back, if they use it outside office hours, and if answers score well against what the organization itself would say.

3

The evidence

The pilot ran in short loops: each change was specified, built with tests and deployed to the pilot. Three of the changes:

ChangeWhat shipped
An easier first two minutesTyped country search, phone formatted while typing, lenient postal codes, a success notice
The interface in the newcomer's language32 interface languages, machine drafts labelled as machine translations, native-speaker review requested by community size
A rebuilt knowledge base10,000+ cleaned pages from 60+ organizations

The Lab shows how this work was ranked by its impact on newcomers and the effort to build it: What to build first.

4

The decision log

No.WhenDecisionWhyWhat followed
D-01Jun 2026Give the pilot its own environment, separate from the main productThe pilot needed regular changes, while the main product followed its own review and release schedule.The pilot stayed in step with the main product.
D-02Jun 2026Moderation before display, failing closedEvery public message is held, classified, shown only if cleared; a human reviews the rest.Still stands.
D-03Sep 2026Languages now, labelled as machine translationsMany users were new to Canada and often not working in English.Machine drafts carry a visible legend until a native speaker has reviewed them.
D-04Oct 2026AI drafts room summaries; a moderator approves themA published summary carries the organization's name.Room summaries are drafted from the last 200 messages and approved by a moderator before they are published.
D-05Oct 2026Unread badges refresh once a minuteThe realtime service bills for live connections, and a one-minute delay is fine for a community chat.Still stands.
D-06Sep–Oct 2026Every paid call capped per person: translation, voice, emailEvery paid action costs money and a runaway caller could drain the budget.In place for translation, voice and email.
  1. 0160+ organizations' websites
  2. 02Crawl, clean, review
  3. 03Knowledge base
  4. 04Chatbot answers only from this
  5. 05Group chat, moderated · support chat
5

The result

Outputs: what shipped

  • 32 interface languages, 5 right to left; message translation on demand into about 130 languages, with a cache and a per-person budget
  • Chatbot conversations with follow-up questions, crisis lines when a reply is blocked, share and email with limits
  • Group chat: search, AI-drafted summaries approved by moderators, unread markers, a replies page
  • Voice input and read-aloud, with budgets; audio kept nowhere
  • A phone layout measured before and after: message area 406 to 681 px on a 375 px screen
  • Three review rounds, each reviewer focused on one area

Outcomes: what changed for people

Piloted
June to November 2026 in Alberta; the results are under study
SettleSMART group chat channels on a phone, grouped by category: housing and daily life, employment and careers, with locked announcement rooms
The group chat on a phone: rooms by category; every message is moderated before it is shown.Fig. 02
6

What I took from it

SettleSMART gave me research experience that is hard to find anywhere in software development.

The research in this job translates into serious UX research: learning what people need from how they use a product, and turning that into product decisions.