Projects
Prediction of Hotspots for COVID-19 and Influenza for the USA
We used publicly available data to predict infections and deaths for each county in the USA for several weeks into the future. We combined primary data about COVID-19 infections and deaths in conjunction with data on the demographic characteristics of counties, the mobility and masking of the population, and the developing understanding of the virus’ incubation. After dimensional reduction of the time-varying and time-independent features to remove inter-dependencies, a Bayesian time-series methodology was implemented based on time-delayed regressors. A similar process was implemented for influenza at the state level in the USA using multi-year infection data. Hotspots could be efficiently and reliably identified in both cases as zones where morbidity and mortality were expected to increase significantly quicker than the national average. The results were provided to government agencies and health systems to help set policy and allocate resources, while also being published publicly for individuals to evaluate their personal risk.
Short-term Demand and Supply Forecasting for Bike Taxis
Bike taxis have become a popular transport option, nowhere more so than in Indonesia where consolidating bike taxis into an app-based platform resulted in the first tech unicorn in the country. However, predicting the fluctuating demand and corresponding in a large city is challenging - especially since the desire for bike rides is strongly connected to the weather in tropical regions. A model to forecast the local demand and supply for the upcoming fifteen minutes was requested, based on current and historic data. The problem was divided into segments: the demand was predicted based on historical information, current trend, and weather information; the supply was predicted based on the historical driver tendencies, the current locations of drivers, and the estimated endpoints of currently active rides. Neural networks and Markov chains were employed to arrive at a model with around 90% forecast accuracy.
Modeling Loan Default Risk from Subjective Expert Experience
Our client, a microfinance company that disburses home loans to low-income individuals working in the unorganized sector, required a system to rate the risk of such loans. The risk of default is high because the applicants’ creditworthiness is not documentarily established, and the loan duration is for a considerable period of time. Sales personnel gather a multitude of information, either directly from the applicant or from different sources. This information becomes the primary data source to analyze and establish whether the loan will be approved, based on an estimation of the probability of successful repayment by the applicant. The principles of Aashiyan have decades of experience in underwriting microfinance loans and are able to make considered decisions based on the available data. However such expertise is mostly informal and subjective and cannot be easily translated into policy or automated risk-scoring. Each evaluation is time-consuming and thus makes it expensive for the organization to scale. Additionally, in the absence of a formal mathematical model, new information is harder to incorporate, and the evaluations suffer from a lack of flexibility. Our system formalizes the subjective expertise of the principals and builds a mathematical structure for risk scoring. The model then was integrated into the client’s data collection system and was able to instantly provide risk-scores as the information was entered. Finally, the model is made adaptable to information about loan outcomes and is able to learn from experience.
Products
Estimated Time of Arrival of Public Transport Vehicles to Stops
An application, sponsored by the World Bank Group, was developed to track all public transport vehicles in the city of Kolkata and disseminate their locations to citizens. As part of the implementation, a model was developed to accurately estimate the time of arrival of vehicles to stops, providing citizens to use this information about their choice of travel mode. The model used existing information collected from the buses during their travels and incorporated additional information related to the day of the week, time of day, weather conditions as well as special occasions. The network of routes was divided into route segments and duration was estimated for each segment. The base model was subsequently enhanced by incorporating the time of the most recent vehicles to traverse the segment weighted by recency. The model was incorporated into the application and citizens have enjoyed the benefits of the estimates since 2017.
Radiaide: A Platform for Radiological Image Analysis and Screening
Radiaide is a cloud-based platform for fast and reliable medical image analysis. It supports multiple disease models and imaging modalities and can be used to instantly provide artificial intelligence based analysis of those images along. It is intended to be used in cases where imaging can be performed in remote regions with instant screening and subsequent validation by remote radiologists. The platform supports state-of-the-art security and privacy features along with an efficient picture archiving and communication system as well as annotative viewers for different images. A series of single-purpose Neural Network models are available to be trained and used. Radiologists and other stakeholders are presented with the model results which can be corrected by authorized personnel. Currently, the platform supports Chest X-Ray images for Tuberculosis screening.