Myna Bolo (‘Speak’ in Hindi)

Women in India still face barriers to accessing sexual and reproductive health (SRH) services due to stigma, discrimination, and lack of high-quality options. The ability of these girls and women to protect and foster their own SRH needs is hampered by a huge gap in knowledge about their health, bias among health providers, and social stigma. Over 3.4 million women in India have a desire to avoid pregnancy, of whom 94% are married. 

Moreover, 71% of girls report not even knowing about menstruation before their first period (NFHS 2021). As women get more educated, stay in school for longer, and aspire to go to college and work beyond, there is a dearth of programs focused on them in this critical stage where their trade-offs with other aspects of life start to increase. To compound these challenges, women living in urban slums have less space and privacy than their rural counterparts, which further hinders their ability to take care of SRH concerns.

Most women in India have historically learned about their menstrual cycles and other SRH queries from their peers, who themselves have often learned through rumors, WhatsApp forwards, or misinformation on the internet. This perpetuates lack of knowledge, continued stigma, and the fear of being judged by family-connection health providers. Community members perpetuate taboos and misconceptions about menstruation and SRH.

Fortunately, with COVID pushing even those in poverty online, more women than ever before now have access to their own smartphones. Nevertheless, existing solutions online for SRH information are limited, and do not consider the local context women are coming from. Satisfactory responses to queries from women living in urban slum settings should look very different from responses to queries from women in rural areas. As mentioned previously, omen in urban slums have little space and privacy, and are overloaded by a variety of misinformation from competing sources; comparatively, women in rural areas have less access to information overall and are less exposed to newer norms.

Hence, women need an unbiased, factual source of information about their SRH that is non-judgemental, confidential, accurate, and actionable given their social norms and existing context. The advice also needs to be personalized for them given their specific conditions with references to local female doctors for prescriptions and medical treatments. Further, how information is communicated is just as important as what is conveyed if we want to make sure they actually understand and benefit from the advice. This is possible via the use of a chatbot through women’s smartphones where they can discuss their problems via text or audio in their own local language.

Example Chatbot experience for the PoC

Correcting misconceptions: 

When the following was entered into GPT (Model 4), 

"kisi kisiko utras me pain ke karan infection hota hai  injection ki waja se TB ki bimari horahi hai ,uSse bhi pregnancy covcive nahi horaha hai" which has a lot of errors in the query. Essentially she means to say that "some people think that due to uterus pain they get infections and these infections can cause TB due to which they cannot conceive a child". She is trying to check if this is factually correct, which it is not. So the response from the doctor/AI chatbot should be trying to correct her misconception in a series of conversation statements. Ultimately, the goal of the PoC is to generate this conversational flow with the help of the training datasets.” 

Then, it gives the following:

Patient:

"kisi kisiko utras me pain ke karan infection hota hai injection ki waja se TB ki bimari horahi hai, uSse bhi pregnancy covcive nahi horaha hai"

AI Chatbot:

Response 1: "Aapke sawal ko samajhne ki koshish kar raha hoon. Kya aap yeh kehna chahti hain ki utras me dard hone se infection hota hai, aur woh injection ki wajah se TB bhi ho sakta hai, jisse pregnancy nahi ho rahi hai?"

Patient:

"Haan, yahi toh suna hai mujhe. Kya yeh sahi hai?"

AI Chatbot:

Response 2: "Mujhe samajh aata hai ki aap kyun pareshan hain. Lekin yeh jo aapne suna hai, woh puri tarah se sahi nahi hai. Uterus me dard hone se TB nahi hota, aur TB se pregnancy nahi conceive karne ki koi sidhi wajah nahi hoti. Kya main aapko is vishay pe thoda aur vistaar se samjhaun?"

Patient:

"Haan, please bataiye. Main bahut pareshan hoon."

AI Chatbot:

Response 3: "Bilkul, main aapko samjhati hoon. Uterus ka dard kai karanon se ho sakta hai, jaise ki fibroids ya endometriosis, lekin yeh TB ya infertility ka karan nahi hota. TB ka sankraman sirdard, khansi, aur bukhar jaise lakshan deta hai, aur yeh typically lungs ko prabhavit karta hai. Pregnancy mein dikkat aana alag karanon se ho sakta hai, jaise ki ovulation ki samasya ya fallopian tubes ka block hona. Aapko agar aise lakshan dikh rahe hain, toh ek healthcare provider se sampark karna behtar hoga, jo aapko sahi jaankari aur treatment de sakta hai. Kya main aur kuch madad kar sakta hoon?"

Patient:

"Nahi, dhanyawaad. Aapne meri bahut madad ki."

AI Chatbot:

Response 4: "Aapka swagat hai! Agar aapko baad mein koi aur sawal ya pareshani ho, toh please mujhse sampark karein. Khud ko dhyan rakhein, aur swasth rahein."

This conversation exemplifies how the AI chatbot can gradually and empathetically correct a misconception by asking clarifying questions, providing factual information, and offering further assistance. It maintains a supportive and non-judgmental tone, ensuring that the patient feels comfortable and understood. The conversational flow illustrates the key objectives of the PoC, demonstrating the chatbot's ability to handle complex, error-filled queries and respond with medically accurate and compassionate guidance.

Myna Bolo focuses specifically on women and girls in Indian urban slums between 18-30 years who have at least completed education until Grade 10 (middle educated), have access to smartphones and lack access to sexual and reproductive health resources. Over the last eight years, Myna Mahila Foundation is on track to support 1.5 million such women in urban slums with their SRH concerns, and these experiences have built our understanding for the types of issues women are not able to resolve within the existing ecosystem. Most relevantly, we implemented the Myna Health App, which included sections for period tracking, short videos, SRH daily tips, and public & private chats.

Building on the work of Myna Mahila, we now propose to build a SRH chatbot, Myna Bolo (“Myna Speak” in the imperative in Hindi), which will integrate LLM technology into the Myna Health App through employing a competent LLM model. This tool will empower women in India with on-demand factually accurate information regarding their SRH concerns that, critically, is actually relevant for their context. We aim to engineer a solution for a woman to ask questions about her health and get tailored responses for her particular situation after a series of questions and steps. For example, with the existing Myna Health App, one of the existing 10,000 users messaged the Myna team through the private chat (in Hinglish) “I’m having a problem with white discharge.” With the proposed solution, our interactive tool would engage her immediately for further questioning and get her the answers she needed. Given the urgency (real or perceived) of these queries, the limited time a woman has on her shared smartphone before it’s commandeered by her husband, and the private nature of these questions, it is essential to respond to a query right away to ensure that the woman sticks with the conversation and receives accurate information. It wasn’t possible before to respond within such short windows in a scalable way, but now, thanks to the use of novel conversational LLM technology, Myna Bolo can solve this problem.

To develop the Myna Bolo LLM-based chatbot for women, we will be working with the following teams with their roles and responsibilities listed below. 

1. Community workers 

2. External Technology vendor 

3. LLM Experts team

4. Internal Management team 

Community Workers

Myna Mahila has an existing initiative, Rani Work, that employs local women to create training datasets in local dialects in urban slums. Currently, we have 1,200 Rani workers, speaking 15+ local languages with minimum 10th grade education level (and 25% with college graduate degrees), and these workers will  create these datasets for the proposed work. Further, the Myna Health App already contains data from women conversing with local doctors and the Myna team privately. 

External Technology Vendors

This vendor will focus on model training, knowledge base restriction (to SRH - specifically family planning in the first phase), embedding addition, integration with mobile application, performance optimization, security and compliance,  quality assurance and testing, maintenance and support, documentation and training the team for operations and management of the chatbot.  

LLM/Glific Team (Tech4Dev)

This team will focus on the prototype development using our data to train the model on the Glific platform. This will ensure that the training process yields a model that accurately understands and responds to queries related to women's SRH. The other focus includes quality assurance by testing the prototype for accuracy, reliability, and usability; Oversight of the external vendor and ensuring they meet the milestones. As LLM experts they can provide technical training and advice/ consultation. Lastly, they will review and validate technical documentation provided by vendors. 

Internal Team 

The internal team includes: 

• Overall Project PI

• Project Strategy Lead

• Technical Product Manager

• Content Generation Lead (Queries)

• Content Generation Lead (Medical Responses)

• Software Developer

Furthermore, 70% of the staff at Myna Mahila Foundation are from our beneficiaries since 2016 and the senior management team is led by women. 

We would utilize this grant to hire a project manager.

Currently, the Myna M-Health Application has over 21,000 downloads. 

Myna Bolo is a one-stop solution for women on their phones integrated in Myna’s M-Health application. The features include:

•  Tailored conversations: Provides accurate, natural and empathetic responses to users in colloquial Hinglish language

• LLM Integration: Uses intent detection, contextual prompts and vector search in GPT-4 and guides users to SRH video topics and period tracking

• Telehealth referrals: Allows users to schedule appointments with healthcare professionals

As of September 2023, our proof of concept is: 

1. Trained to tackle real questions over 17,000+ queries on SRH generated by Myna Mahila's "Rani Workers" (previous MIT Solve winner) from local communities used as a training dataset to improve LLM responses. 

2. Tailored for Local Terminology: Local slang, metaphors, and hidden meanings have been documented to create local term guide 

3. Verified by Network of Doctors: Over 4,300 reviews by network of local doctors collected to verify the medical accuracy and relevance of Myna Bolo responses 

As of Sept 2023, results from our first pilots on the Proof-of-Concept reveal: 

• 70% user evaluation of Myna Bolo (rating by women from the end-user community as of Sept 2023) indicates that the bot is helpful but requires additional fine-tuning through further development of our localized LLM technology 

• 94% doctor evaluation of Myna Bolo (rating by a network of doctors in and around Maharashtra as of Sept 2023) conveys that the responses are medically accurate, empathetic and relevant to user queries. 

Our model seeks to continue to debias Al through iterative testing by Myna Mahila's trained "Rani Workers" and dedicated network of doctors.

PATH TO IMPACT 

Since Myna Bolo specifically addresses an existing gap in product and delivery. Sexual and reproductive health requires more training data from diverse populations of women and living in diverse and more remote geographies, Myna Bolo has several unique advantages that make AI more equitable. These populations are traditionally underrepresented online, and the proposed solution generates new training datasets through women located in similar communities as those for whom the datasets and chatbot are to be used. Further, the Rani workers will help in the creation of the Myna Bolo chatbot from its design, usability, ease of use, etc.

If Myna Bolo is successful, we will make it open access and share it with a variety of NGOs, government organizations, and any other stakeholders looking to empower their beneficiaries with critical information pertaining to SRH, including non-judgemental personalized solutions. It would be extremely scalable and straightforward to update the tool with more languages, as well as with additional information tailored even more granularly to specific slums. In the future, this localized SRH chatbot will also be tailored for rural areas, and will be expanded to support women in other domains as well, such as with information related to livelihood and career development. We could follow the same set of four steps needed to tailor the application to the livelihood domain to provide women with personalized encouragement and locate local jobs available to match their needs.

Myna Bolo is promoting SDG-3 of Good Health and Wellbeing and SDG-5 of Gender Equality as women need an unbiased, factual source of information about their SRH that is non-judgemental, confidential, accurate, and actionable given their social norms and existing context. The advice also needs to be personalized for them given their specific conditions with references to local female doctors for prescriptions and medical treatments. Further, how information is communicated is just as important as what is conveyed if we want to make sure they actually understand and benefit from the advice. This is possible via the use of a chatbot through women’s smartphones where they can discuss their problems via text or audio in their own local language. The Myna Bolo chatbot will serve this purpose and increase access to SRH information in a localized context. 

In the longer term, we hope to make the chatbot open access and share it with a variety of NGOs, government organizations, and any other stakeholders looking to empower their beneficiaries with critical information pertaining to SRH, including non-judgemental personalized solutions.

We have piloted a number of evaluation tools at every step to examine and document the decision making process against the milestones. Technical milestones are evaluated against success metrics. In the second set of user testing criteria where we employ super users from the community side and medical professionals, we’re using behavioral science methods to get feedback from them and implement these changes. 

Similarly we also undertake qualitative research where we have community members test out the bot on their phones and during the process, we collect feedback to study impact and create feedback loops between the field, program teams and tech partners. 

Apart from a 60-person Myna team, 10 full-time members: Founder/ED, Program Manager, Data Associate, two Quality Assurance Supervisors, Accounts and Payments Executive, and four field coordinators to support the community workers. 

The technology vendor is contracted out to an external agency.

The Myna M-Health Application was first developed as a web application in 2020. The Myna Bolo chatbot was integrated into the application in 2023. 

It is crucial that our team members complement each other well and bring a variety of skill sets and perspectives to the table. We ensure diversity from top down in our cultural and economic differences. Further, we encourage perspectives to be heard in bottom up order. During team debriefs, people working most closely with women from the communities who are generating the queries speak first and they are given the most air time. We constantly relate back to our values during the meetings every time there seems to be diversions, and we regularly engage in open and honest dialogue (including mistakes/failure). If our team believes in diversity and inclusivity themselves, we would be equitable in recruiting women of different ethnicities and cultures for Myna Bolo. We provide training to be unbiased in the recruitment process to ensure we are not discriminating. For instance, our Muslim coordinators visit Hindu homes to speak about Myna Bolo, and vice versa. When our diverse team talks to each other so openly and freely about their lives, others around them tend to open up as well. 

Finally, we need to invest a lot more in leadership development of our team members, providing them with mentorship to grow more confident of themselves and their team. 

Notes: While we want to maintain a team of majority women (since it resembles our beneficiary base), we need a few more male members on the field who can help drive more credibility about the platform among husbands and community leaders. We also want to maintain our balance of field and office based team members, where the former executes all plans on the field and the latter works closely with the field team to ensure strategy, operations, budgeting, and resource management are going as planned. We need to continue growing our team who focuses on software development and data science experience who can help us continue building our AI-generated chatbot.

Myna Mahila, through its data-driven and community-oriented approach, is poised to create a transformative impact on the lives of women in underserved communities.The women we serve come from backgrounds where online engagement is limited, and the digital landscape has little insight into their specific language, behaviors, or preferences. However, the advent of generative AI technology presents an exciting opportunity to bridge this gap and empower these women with essential knowledge and support.

Our innovative approach establishes a powerful feedback loop, empowering local women in our RANI workforce to actively participate in the process. These RANI workers will receive training to contribute local healthcare-related queries and responses, creating a valuable dataset encompassing misconceptions, local terminologies, and a wide range of health queries, from basic to complex. This data, in turn, is used to fine-tune a Large Language Model (LLM) that is integrated in Myna Bolo. 

The key strengths of this approach are its localization, absence of bias, and cultural relevance. The information provided is tailored specifically for women, by women, addressing their unique needs and circumstances and also providing a source of income to the RANI workers.

Explain how you will bring in money to fund your work, whether through sustained donations and grants, selling products or services, service contracts to governments, raising investment capital, or a combination of all. In the long term, your revenue streams should cover your expected expenses. To see some examples, take a look at these common revenue models.

Maximum 500 words

In the short term, our financial sustainability plan centers around securing grants that align with our mission of empowering underserved women with AI-driven healthcare knowledge to support the pilot phase and diversify community testing.

In the long term, our vision for financial sustainability involves strategic collaborations with government agencies and prominent non-governmental organizations (NGOs). These partnerships will open doors to reaching a larger user base and expanding the reach of our services. As part of this long-term strategy, we plan to introduce operational costs.

Operational costs will encompass essential elements, including technology support, field team salaries, and program maintenance. By charging these operational costs, we aim to establish a self-sustaining model that covers the ongoing expenses associated with our initiative.

Through this combination of short-term grant funding and long-term partnerships with government and other NGOs, we want to secure the necessary financial resources to implement the solution.