Taxi Mobile SolutionTMS Embarks On AI Journey To Robust Its Taxi Booking Platform

In the Taxi business, AI has helped immensely in understanding customers’ needs, their route preferences, spending habits, and communication patterns to solve their commuting issues. On top of it, it has made the passenger experience safer and more customized than ever before. TMS wanted to accelerate the development of its app on these lines to provide its clients- taxi/commercial fleet owners a robust taxi booking platform. The development envisaged the requirement of training data based on both Computer Vision and Natural Language Processing as per the client use case requirements. We helped them power driver monitoring systems and driver-passenger chat solutions with accurate training and data required for their ML models.

Driver monitoring systems in the market are still in their infancy. While performance and safety validation is paramount, there is a significant gap when it comes to effectively recognizing the complex and nuanced cognitive and emotional states of the driver. TMS encountered a major challenge of keeping up with the pace and cost of labelling its client imagery and conversational data which was in sheer volume and variety. There is no one-size-fits-all solution that meets all the data labelling needs. The robustness and effectiveness of the algorithms depend on the availability of high-quality driver activity datasets. The challenge was to have a dedicated, highly skilled workforce continuously working on the job and assisting with client-specific data annotation.

The team helped TMS develop high-quality data sets with a feedback intensive communication structure. Two teams of 8 members each for imagery and video annotation, and one team of 5 members for tagging text, labelled vast numbers of CV and NLP datasets that TMS needed across its client use case requirements. Our labelling teams made focused shifts to keep business rules intact as edge cases came up, and adjusted taxonomy and quality rubrics to reflect the current training data objectives. We designed the labelling workflows to optimize accuracy and simplified the number of labels per image/video frame to eliminate errors.