Ankur Jain, the former chief product officer of BharatPe, left the fintech giant last year to venture into AI and healthcare. Together with G V Sanjay Reddy, the chairman of Reddy Ventures, he co-founded Jivi.ai, which happens to be Jain’s second stint in entrepreneurship.
Jain, who has a background in AI and machine learning, founded Jivi in January 2024. Within months of incorporation, the startup announced its first model Jivi MedX, which outperformed popular models like Google’s Med-PaLM 2 and OpenAI’s GPT-4 on the Open Medical LLM Leaderboard.
In an exclusive interaction with AIM, Jain revealed that the company plans to launch a series of models focussed on healthcare in the coming months.
“While Jivi MedX is a text model, we are working on a series of other models that we internally call a model cluster. For example, there could be a different model that specialises in diabetes, and for ophthalmology, there will be a different model.
“The next model from Jivi will be a vision model. We are working on a multimodal MedX,” Jain said.
The vision Jain has with his new startup is to build an AI medical companion that the eight billion-strong population in the world can use for free. AI is finding use cases in other domains of healthcare such as drug discovery and genoming. However, Jain’s startup is focussed on primary healthcare.
Not Just a Chatbot
The startup’s aim is to create a product which will be useful to end users as well as doctors. Jain revealed that the AI medical companion he is building will have voice capabilities built in, which means everyone would be able to converse with the model in their own languages.
“We will have voice capabilities and the models will eventually understand the top 25% of the world languages, which covers around 90% of the population. However, the technology is not there yet. Even OpenAI has not released the voice capabilities for GPT-4o,” the Stanford alumni pointed out.
However, it’s just a matter of time, according to Jain, who comes from a family of doctors. “I think in six months, it will become mainstream. We are not building a chatbot, people will have to explain their symptoms to the model and for this, voice is the best medium.”
Regulatory Challenges
Jain’s endeavour is bold since healthcare is one of the most highly regulated industries in the world. Moreover, regulators around the world are contemplating regulating the technology.
“For instance, we already have telemedicine regulation, whether you use AI in telemedicine or not, telemedicine rules will apply. In the end, the healthcare side will always supersede the technology, whether AI or non AI, we have to abide by the regulation,” Jain said.
Nonetheless, Jain acknowledges that governments worldwide are contemplating regulations and changes are imminent.
“We already adhere to standards like ISO 42,001, and we must comply with evolving regulatory frameworks, not only in India but globally.
Similarly, in AI, especially in healthcare, we welcome and adhere to regulations. Regulations ensure that only serious players can enter the market, driving the delivery of quality products to consumers,” Jain added.
Powered by Llama 3
Jain revealed that JiviMedX is powered by the Llama 3 8 billion and 70 billion parameter model. The business started by trying out the existing models, but none of them produced good enough results for the medical domain.
“We tried Mistral, Llama 2 and a few other domain-specific models but none of them helped us in achieving the results we were looking for. The accuracy at best was around 55-60%, which was not good enough,” Jain revealed.
To make the models more accurate, the startup experimented with various approaches initially, including direct preference optimisation (DPO) and reinforcement learning with human feedback (RHLF) frameworks but these methods did not yield the desired results.
“So we settled with odds ratio preference optimisation (ORPO). Interestingly, not many people in the world have used it so far. We kept experimenting with it, akin to tuning hyperparameters. We conducted approximately 150 experiments to achieve the desired outcome,” Jain said.
Currently, JiviMedX has an average score of 91.65% across Leaderboard’s nine benchmark categories. However, Jain adds that even 92% is not good enough and efforts are underway to increase the model’s accuracy level to 95%.
While MedX uses Llama 3 for its intelligence, the other models Jivi is building will not necessarily be powered by Llama 3. Other models in the cluster could use a different model for intelligence depending on the use case.
‘My First Hire was a Doctor’
“We plan to launch a product in the coming months, but we need to be 100% sure that the model works and does not hallucinate. Currently, the medical community is evaluating it and we will launch it as soon as we get the green light,” Jain said.
Interestingly, Jain revealed that even though it’s an AI startup, his first hire was a doctor. Currently, his team consists of 24 people, and nearly 40% of them are doctors.
“We are involving doctors from Day 0 basically. Half of my team is full of in-house surgeons and physicians. We have also collaborated with the doctor’s community throughout. We built a prototype, shared it with the community and incorporated their feedback, that’s the process,” he added.
Jivi’s Secret Sauce
Jivi trained MedX using domain-specific data gathered from web scraping, along with information sourced from books, medical journals, research papers, and clinical notes. However, the data alone does not solve the problem.
“An article about diabetes found on the web tends to be quite generic, typically targeting end users. In contrast, our analysis delves into the past 50 years of diabetic research data. We aggregate insights from various sources, compiling them into a comprehensive knowledge base. This is our secret sauce,” Jain pointed out.
Another challenge the company faces is keeping the data up to date. For instance, if there is new research in diabetes, the startup has to ingest the data into the model.
“So, we have to ingest every new development into the model every month. We plan to do this every week. Currently, we are ingesting once a month,” Jain said.
Can AI Make Money?
Creating an AI model is one thing; monetising it is a completely different challenge. AI requires huge investment, but building a sustainable and profitable business model remains a challenge.
Jain plans to release the product for general consumers for free. However, the startup also plans to create a B2B segment where hospitals, pharma companies, and insurance providers will provide Jivi’s product to their end users as a software-as-a-service (SaaS) product.
“Our business model involves selling premium services and also cross-selling. We can connect end users to diagnostic services, and we can also cross-sell e-commerce products like medical devices,” Jain said.
