ICTworks

Generative AI is at the top of the hype cycle with everyone claiming they have the first or best or newest use in international development. While there is much froth, there are social impact organizations that provide dynamic and personalized advice on using AI, based on years of experience with machine learning.

One of those firms is Turn.io, which spun out of Reach Digital Health (previously Praekelt.org) after years of working with AI-powered solutions, especially chatbots. Turn.io helps social impact organizations integrate directly with the WhatsApp Business Platform and scale effective support to people everywhere.

I’ve always admired their work on safe & responsible large language models (LLM) in chat for impact services. Here are three benefits they’ve found from incorporating LLM chatbots into your service delivery model:

• LLM services are available 24/7 without a coffee break or holiday and never get tired or angry, even when users can be grumpy or outright mad.
• LLMs provide automatic bidirectional translations between users and human operators when operating across time zone or language barriers.
• LLMs automate routine content queries and help route hard questions to designated human experts.

There are still billions of people globally needing essential health and education. When LLMs are deployed at scale, they will be an incredibly cost-effective means to have an impact.

LLMs in Chat for Impact Services

LLMs also have the potential to be transformative in how chat for impact services are designed. Autocomplete patterns have been commonplace for years in word processors and other text-authoring environments. Similarly, LLMs can assist program managers to design services for impact, and continuously improve the process. LLMs can constantly monitor, evaluate, and provide feedback as a service is used.

Chat for impact services designers can utilize feedback to iteratively improve their service, especially responsible LLM deployment. Constant feedback enables organizations to test different mechanisms for implementing LLMs whilst ensuring transparency, fairness, and accountability.

Turn.io is now helping organizations enhance their chat dialogue capabilities by integrating with ChatGPT and using a custom Playbook to guide organizations on connecting with OpenAI to retrieve responses and deliver them to users on WhatsApp. Turn.io is also helping organizations leverage machine learning models hosted on Huggingface, where organizations can enhance the meaning inference of user requests within the Turn.io platform.

2 Real World GenAI Chatbot Examples

Enough on how Turn.io impresses me. Let’s hear about real-world examples from health and education that can impress you.

GenAI Chatbots in Education

Rising Academies is addressing the maths education crisis in sub-Saharan Africa using ‘Rori‘, a chat-based virtual maths tutor that operates on affordable mobile phones in areas with limited connectivity. They developed hundreds of conversational lessons aligned to standards from UNESCO’s Global Proficiency Framework for Mathematics, which, via a mix of expert-written scripts and LLM-powered generative chats, help students improve their maths.

Generative AI allows Rori to have four benefits that are unique to artificial intelligence solutions:

1. Rori is scalable.

Initially developed as a response to COVID-19-related school closures, it’s designed to assist millions of students globally who lack access to advanced technology or inexpensive internet. Rori functions effectively on commonplace messaging apps like WhatsApp and already, tens of thousands of learners are utilizing Rori in Sub-Saharan Africa, yielding promising results.

2. Rori can dynamically process students’ answers.

It can also recognize a variety of formats, including different ways of expressing fractions. Whether a student types “3 over 8,” “3/8,” or “I am not sure, maybe 3 ovr 8?”, Rori can understand and generate accurate and contextually appropriate feedback. This flexibility ensures that students receive the support they need, regardless of how they present their answers.

3. Rori can identify students’ intents and preferences.

By analyzing students’ requests and expressions, Rori can understand if a student requires more challenging exercises, manageable tasks, or hints. For example, if students mention feeling tired, Rori can recognize this intent and suggest taking a break, ensuring students’ well-being is prioritized.

4. Rori can answer more complex and open-ended conversations

It can converse on a range of topics, including metacognitive skills, historical maths figures in Africa, and real-world applications of Maths. The LLM starts with a skeleton for the conversation, as well as guardrails, but each conversation is unique and depends on the students’ questions.  Try it yourself – message “hi” to +1202982447 on WhatsApp.

These free-form discussions are not with unmonitored students yet, due to the risk of opening pathways for inappropriate conversations with young learners. Rising Academies has safety and quality assurance measures in place and deeply tests generative chat before releasing it to all users.

GenAI Chatbots in Healthcare

Guidepost is dedicated to improving the quality of life for individuals with diabetes with advanced AI clinical decision-support systems, automated WhatsApp chatbot patient communications, and clinician training infrastructure.

Their coaching services help people with diabetes to lead healthier and longer lives. Guidepost’s new WhatsApp chatbots have driven significant growth – up 89% in 2023 – and almost half of the patient base is on a chatbot-initiated program.

Guidepost began implementing an “Auto-coach” platform in October 2022, with the aim to automate 80% of coaching tasks using AI technology. Initial uptake has exceeded expectations, with 33% of patients becoming high engagers. Chatbots help patients set meaningful, achievable goals, submit glucose- and other readings and track progress.

Anecdotal feedback includes phrases like, “it didn’t seem like I was engaging with a chatbot” and “I realised my overall diabetic control had improved!”.

Since mid-2020, Guidepost platform capabilities have been extended to incorporate specific tools such as AI-based trend identification, projected HbA1c estimation, and AI-enabled suggestions for clinical interventions, all of these contributing to Guidepost’s effective follow-up and issue escalation planning. Feedback from human clinicians on AI-generated information across thousands of interventions helps Guidepost build stronger and more unique AI models.

Guidepost’s comprehensive approach, supported by AI, empowers individuals with diabetes and the clinicians that serve them to manage their condition and improve their well-being.

A special thanks to Lieze Langford for helping me with this post.