We tease the problem out of the system and use AI to resolve it: Shekar Sivasubramanian, CEO, Wadhwani AI

Shekar Sivasubramanian spoke to indianexpress.com on how AI is being used to solve problems in agriculture, health and education sectors, the challenges of using AI, and the ecosystem needed to nurture it. Edited excerpts:

Venkatesh Kannaiah: How do you pick the areas/themes to work on? Is it based on problem statements from partners?

Shekar Sivasubramanian: We primarily work in three domains which have a huge social impact: health care, education, and agriculture. We make sure that we are deeply embedded in these domains. When it comes to AI, problem identification is the most critical step.

These problem areas are not obvious or available upfront. They need to be teased out of the system. The definition of the problem requires the intersection of domain knowledge, understanding of the Indian technology space, and an appreciation of the entire ecosystem that will use the solution. We build the problem themes based on a holistic understanding of partner requirements and the feedback from the field.

After identifying the partner, whether it be a government or a nonprofit, we have a deep dialogue to define the contours of the problem. Obviously, it would need to have a substantial social impact. The details of placement of AI could be a new application, or an addition to existing technology.

A variety of stakeholders use our AI solutions. Cotton growing farmers, frontline workers in the health segment, doctors, teachers and students in the education sector, and officials who monitor these issues on a day-to-day basis.

Venkatesh Kannaiah: Can you tell us more about your AI-powered pest management solution, and the impact it has had?

Shekar Sivasubramanian: Cotton is a large crop in India and one of the challenges is the appearance of a specific pest called pink bollworm. The agriculture units of the government have set up traps for catching these pests, based on best practices advised by agricultural research in India. After catching these pests, the farmer may need to take it to a government unit where it might take a few days to find out the nature and intensity of the infestation.

Our solution is simple. We ask the farmer to take a photo of the trap contents on a contrasting surface, upload it on our app, and it takes a few seconds for us to advise the farmer. If the number of pests are high, we give the farmer a generic pesticide name, and provide them the overall context of how they should treat the problem. It saves the farmer time and money. Alternatively, we advise them to wait and watch, or tell them that there is a hazard of infestation. We created an AI-backed ecosystem to help people understand that such models are feasible.

We have educated over one lakh farmers about this pest management tool, which is now being embedded in the National Pest Surveillance System (NPSS), where it is likely to grow tenfold in the next couple of years.

Venkatesh Kannaiah: What are the other AI applications you are working on in the field of agriculture?

Shekar Sivasubramanian: Whatever work we are doing with cotton pests, we want to extend it to different crops, and all of it will lead to NPSS. We are also working on land mapping and identifying what crops are growing where using satellite imagery. We are also conceptualising a national-level spectral library in partnership with the Mahalanobis National Crop Forecast Centre, Delhi. This is in its early stages. There is another interesting one called the Integrated Agriculture News Monitoring System (IANM).

The IANM, an AI tool, reads thousands of newspaper articles in various Indian languages, identifies trends and issues of importance to agriculture departments in the government and alerts them within an hour. There are 14-15 government departments which are being alerted. The issues range from pesticides, plant protection, and pricing of agri commodities to weather, productivity, water availability, drought etc, and these government departments can initiate action immediately instead of waiting for reports to come from the ground. We can monitor social media too, but we are not sure of a certain level of authenticity, and hence we do not have it in our reporting process.

Venkatesh Kannaiah: How do your health solutions, like event-based disease surveillance work and what are your AI solutions with regard to tuberculosis?

Shekar Sivasubramanian: We have built an Integrated Disease Surveillance System based on tracking information through newspaper reports. Earlier, the government would get such information based on reports from state offices . This is a time consuming and laborious process. Our solution can detect infections and disease patterns based on health events reported across mainstream and regional media, and report it on the same day, allowing greater time and control for the government to identify outbreaks and mobilise their machinery for controlling such outbreaks.

We have a unique solution for tuberculosis tracking too. As soon as a patient comes in, registers and gives his basic data, we provide a score of 1 to 100 on adhering to the medication schedule for tuberculosis. Now, if a person does not adhere to a regimen, that person is likely to spread the disease, likely to become multi-drug resistant, and in due course, going to be a bigger burden on the public health system. Based on large amounts of previously-collected data on patient adherence to the TB medication regimen, we can predict whether the particular patient would adhere to the regimen or not; doctors and public health officials can follow up with such patients, monitor them more closely rather than allow for the problem to fester on.

Venkatesh Kannaiah: Can you tell us about your other interventions in the health domain using AI?

Shekar Sivasubramanian: One of our AI solutions is a Clinical Decision Support System. It employs a data collection system where before the patient meets the doctor all basic parameters, symptoms and other relevant information is collected. One, it saves the time of the doctor, who could spend his time on diagnosis, rather than data collection. Secondly it is stored in a database, and thirdly, the system also throws up some diagnostic suggestions based on the symptoms and the data collected. The tool does a professional diagnosis and a prediction on the top three diseases or the issues that the patient is facing right now. The doctor can use this as a diagnosis as a clinical decision support system and enhance his utility.

For instance, Rajasthan and surrounding areas report cases of silicosis. Our solution can analyse the x-ray of the person and identify if the person has the disease or not. It is like a screening tool, and in medicine screening tools have to be well developed to build more advanced AI-based diagnostic tools.

We have another digital tool designed for reading TB test strips used to identify drug-resistant strains. In this process, the sputum of a TB patient undergoes a chemical reaction, and the response is reflected on a strip of paper resembling a barcode. Interpreting the results can be challenging. Our AI tool reads the strips and provides accurate results with a precision of 99 per cent.

Venkatesh Kannaiah: Can you throw light on your AI Centres of Excellence at various government departments. How does it work and what is the impact?

Shekar Sivasubramanian: We have established these AI Centres of Excellence at various ministries to facilitate a dialogue and to build tools based on their requirements. We also train and help people to understand the potential of AI solutions, enabling them to identify problems where AI can be used.

Venkatesh Kannaiah: What are the challenges and opportunities of using AI for social impact in India?

Shekar Sivasubramanian: As an organisation, we have understood the need to have a sustained presence across states to collaborate on using AI for social impact. We need to identify the problem statements and the specific operational data and work on it till we get it right. The nature of AI is not going to be generic. It is very domain specific. It is not a foundational technology like UPI.

Domains are important in AI. For reading an x-ray, we need to integrate large amounts of diagnostic data into the system and then read the x-ray. We need to be constantly upgrading ourselves on the innovations that are taking place in the space, at the community level, the individual level, the organ level, and even at the cell level to improve our reading, and to enhance the utility of our AI solution.

We have barely touched the tip of the iceberg in understanding how AI can be used for social impact. There are hundreds of implementations of AI that you can look at across the spectrum. We need to be systematic for it to work on scale. The mantra is always: identify the problem, get data, do the solution, make sure it is responsible, deploy it early, get feedback, iterate the model, get it performing well, alert people where it will go wrong, make sure it integrates into an ecosystem, and finally make it work.

Using a set of implemented solutions, we need to establish higher level connections of how decisions can be enhanced using AI, on the basis of already deployed solutions. This is quite a challenge, but this is how it needs to be implemented.