Late last year, e27 reported that the Singapore government had a Facebook chatbot in the works. The chatbot would be used to push out regular news alerts and government announcements from the government information portal Gov.sg to subscribers.

Now that the messenger bot is in operation, its developer, San Francisco- and Singapore-based AI company KeyReply, has published a Medium post detailing the bot’s development cycle: the planning phase, building process, testing procedures and the features that were implemented.

It’s quite a lengthy read, but here are some key points summarised from the post.

The planning phase

The directive from the Ministry of Communications and Information (MCI) was that the bot had to help citizens consume and access information more easily. Beyond that, KeyReply had a lot of leeway to experiment and test out different ideas.

In a separate email interview with e27, KeyReply’s CEO and founder Spencer Yang elaborated on what some of the initial ideas were.

“Our initial ideas for the bot focussed more on ‘thematic issues’, which we later renamed ‘current affairs’. For example, we wanted to make some of these themes more prevalent and create more content around these issues specifically. However, as the discussions progressed, we moved towards making the bot more widely applicable and equipped it with more abilities to answer general questions,” said Yang.

Also Read: From chatbots to intelligent things: Here are 5 exciting industries startups can focus on this year

Besides tailoring the scope for mass consumption, the team also had to design the bot’s personality to appeal to both foreigners and citizens; functionality is important but making the bot approachable so that users would feel comfortable engaging with it is also crucial.

“A few different ideas we had for tones and personalities were ‘strictly business-like’, ‘snarky but informative’, and other such funny ones, but later we stuck to this ‘friendly and a little funny when you ask it non-governmental related questions’ personality,” said Yang.

Then came to the bot’s dialogue options. The team would “plot out all these functionalities and dialogue branches in a map and write out the initial cut of what the bot would say.”

In other words, the team had to create a detailed mock-up of bot’s workflow and functionalities and understand thoroughly how it was going to act in different situations.

Beyond that, the team also made sure that the infrastructure used to host the bot was both secure and scalable. To the end, the team selected Microsoft’s cloud computing service Azure as the bot’s platform.

Also Read: AI-assisted shopping: A chatbot is now available in major Singapore shopping malls

“Azure provides enterprise grade service-level-agreement (SLAs) on services, 24/7 support, and round-the-clock service health monitoring. It is easy for us to setup machine learning pipelines in Azure Cloud, and it also has great developer tooling for building and hosting modern Node.js (a JavaScript open-source, cross-platform runtime environment) web applications,” said Yang.

Building process

Now comes the hard part: putting these ideas into development. A substantial effort was put into ensuring the UX was smooth and intuitive, and that the correct information could be disseminated quickly – and understood by folks from different background.

Thus, the first step was to segment the users into two groups: those with a specific question, and those who are merely browsing.

Image Credit: KeyReply

Next, users had to, at a glance, be able to navigate to the right channel. For this purpose, menu cards with striking visuals were created.

Image Credit: KeyReply

One unique aspect of KeyReply’s chatbot is its ability to push content to users as opposed to just waiting for them to pull content. Users can subscribe to news, videos, or Facebook feeds from the government on daily, weekly basis or just on weekdays.

Image Credit: KeyReply

For users who would rather rest their eyes, they can choose to listen to the news by activating “a dynamically generated voice clip that reads the news by the page aloud.”

The team also realised that users watching the video may not want to step out of Messenger, so they included an option that would allow viewers to watch the video within the app (albeit a half-screen as opposed to full-screen).

Next comes the challenge of information retrieval. Facebook Messenger is not the most ideal platform to throw in huge chunks of information, so instead, KeyReply’s bot pulls up multiple sources users can click to for in-depth information.

Image Credit: KeyReply

Users do not always have to navigate away from Messenger to retrieve information. For certain requests such as weather updates, microsites are pulled up and displayed within the app.

Image Credit: KeyReply

Behind the scenes

For the administrators of the chatbot, a whole lot goes on behind the scenes to ensure the experience is optimised.

For that purpose, the KeyReply team has a backend dashboard that allows them to fine tune and track content through broadcasting, feedback and analytics that allow them to improve the bot’s functions. This includes improving the accuracy of searches and better personalisation of content by leveraging machine learning and Big Data.

Also Read: 5 Singaporean chatbots you will actually use

Other backend processes that the team is using to improve performance include optimising natural language processing – being able to recognise localised words.

The team also continually conducts rigorous internal testing to ensure every functionality, every step and every dialogue flow runs efficiently.

For a more exhaustive breakdown of the chatbot’s creation story, read the original Medium post here.

—-