This proposal outlines the scope of building the Clara blood donation bot that provides useful information to ARC users on blood donation.

This project focused on building only the backend of the bot .

This project will take 5-6 weeks to implement and cost $47,500 in fees. This includes a 25% discount for Red Cross.

Building the AI system involves three parts -

• Definition of intents

• Creation of a taxonomy from the training data

• Coding the actions based on filling the intents

The team will be staffed with a mix of data scientists and developers from Pirates’s network of startups. Two senior AI engineers and one full stack developer will work on the project and other resources will be pulled in as necessary to complete the project on time.

 Insufficient Training

Many questions such as “What happens to blood I donate ?” and “Can I donate blood if I'm over 65?” come back with a response saying “I'm sorry I don't have any information on the subject.”

Bad Language Model

Questions such as “How much time does it take to donate blood ?” ends up bringing back invalid information suggesting that a bad language model is being used.

Question Processing

Queries such as “Can I get any disease like AIDS or Hepatitis or any other disease by Blood donation ?” are replied with answers like “You cannot donate if you were ever diagnosed with HCV or HBV regardless of treatment or current status.”.This indicates that the questions aren’t being processed correctly

Invisible Flows

There are implemented flows in the conversation flow which aren’t accessible from the chat window. For eg. typing “Cancel Appointment” or “Restart” gives us access to new conversation flows for which button options are missing. We are hoping the user will guess these words and access these flows

Navigation

Many flows lead the user to a dead-end without an option to go back. There are also flows where there is a genuine need for a back or a cancel which is missing. For eg. “Schedule an appointment” --> "What kind of donation would you like to make?" --> "Not sure"

Multiple Responses

Some queries bring back multiple / duplicate responses from the server. For eg. Querying “Blood donation requirements” at any time

Unclear Interaction

Whenever a list of buttons is shown, free form text should be disabled and the user should have only way to interact with the bot. This leads to a lot of bugs where the bot gets confused about the context

Mishandling of Context

Bot is not able to keep a track of the right context during conversation. For eg. Moving on from donation to other topics is not tracked, not logging in on the login dialog makes the bot think the user is logged in

Load Time

Loading the bot initially and bot responses are generally slow. Initial bot load will take time given the current architecture of loading the bot from Orbita in an iFrame. But this can assuaged by loading a JS spinner till the iFrame loads. A reason contributing to responses being slow could be the voice data that is sent with every response.

CORS

The website is being served from redcrossblood.org while the bot is served from arcchat.orbita.cloud. To enable this, Orbita has had to turn off CORS which is not a recommended practice

Plain HTML

The bot UI seems to have been built using plain HTML, CSS and JS. This means we are re-inventing the wheel and building UI components from scratch, which other libraries have already built for us

Reusing components without re-setting

The list component for ex. Is being used to render a list of buttons. This list is not being cleared when it is re-used in another context, and hence the user sees options from a previous query. For ex. “Schedule an appointment” → “Blood” → “10001” → “No” → “Restart”

Data Validation

There is no validation on input data like zip codes, date format, year etc.

No option to enable / disable voice

f we load the bot directly from Orbita using the URL: https://arc-chat.orbita.cloud:8443/oeapi/bot/web we get an option to enable or disable voice and enabled by default. This is probably why voice data is being sent on every response. A similar option is missing on the Blood Donation bot.

Build a new UI in React in the future

React provides us with a lot of reusable components and libraries which will eliminate a lot of UI coding and testing. These components and frameworks provide a slick and modern feel which current users are used to.

Since we are moving to React, it will be too cumbersome to serve it from the current ASP.net backend. Also we cannot have a part of the website in React and have the rest of the website in HTML. It would serve us best to keep the iFrame paradigm and serve the bot from a new node backend.

Incremental improvement of the Language Model

Our language model will evolve over time and at least initially there will be certain intents for which the model will work very well and certain others for which it wont work well. Moving to an ML model will be a staged approach where we will some intents running on a ML model and some intents running on fixed conversation flows