Saarthi.ai automates contact centers using AI agents to help businesses reduce operational costs.Key accomplishments:➡️ Collaborated to develop and maintain two intent classification modules from scratch in 11 Indian languages for Saarthi's multilingual conversational AI product, MAIA, for the debt collection vertical.➡️ Collaborated to implement a sophisticated deep learning named-entity recognition model that identifies dates and times within customer spoken utterances in 11 Indian languages in Tensorflow 2. Instrumental in capturing promise to pay dates and callback dates from the customers, which in turn provided valuable business insights to Saarthi's clients.➡️ Developed a question answering chatbot POC for the Parliament of India to answer questions related to the Lok Sabha and Rajya Sabha. Learnt about QnA systems on the fly and implemented a working POC within 2 days.➡️ Rewrote the model training codebase from Tensorflow 2 to PyTorch Lightning, following clean coding architecture, thereby making it easier to maintain and contribute to. Inference using new codebase is 1.5x faster.➡️ Implemented a more efficient fine-tuning strategy for pre-trained transformer models, thereby reducing training time by 70%➡️ Condensed ~1.1GB fine-tuned transformer models into ~40MB using knowledge distillation, dynamic quantization and TorchScript for deployment, effectively reducing real time inference latency from 1-2 seconds to under 100 milliseconds.➡️ Served deep learning NLU models by developing an API for them, and containerized it using Docker. Single container instance capable of handling at least 200 concurrent requests.➡️ Used Azure Machine Learning services to train and deploy models into production.➡️ Established comprehensive guidelines for maintaining proper branching structures in the team's GitHub repositories.➡️ Collaborated with a team of highly skilled linguists to curate text corpora in 11 Indian languages.➡️ Managed and mentored 3 ML interns.