Aayush Sharma

ConvoZen.AI, an AI-driven contact center intelligence platform. Built ASR system for Indic language calls, achieving a WER below 20% using Wav-2-Vec and Conformor models. Contributed to in-house development of ConvoZen LLM, involving dataset creation, pretraining, and fine-tuning. Instituted an Indic NER model for business entities for call centre conversations, yielding 87% F1 score. Improved Key moments identification framework in calls and Engineered prompts for LLMs for crafting datasets. Worked on model optimizations and deployment, delivering on-demand APIs and real-time processing pipelines.

Prior Framework for Wealth Recommendations. Instituted prior framework using BiLSTMs, Bayesian networks, ARIMA and statistical methods to introduce Advisor Recommendations Affinity module yielding 15% increase in recommendations consumed. Built end-to-end data agnostic system with self evaluation criteria that enhanced problem specific model choice. Implemented framework on use cases such as loan repayment, data anomaly detection and money in motion.

Federated Weak Learning Framework. Built Federated Weak Learning framework to train biocuration models on enterprise domains with unlabeled data. Created Weak Supervision pipeline for NER with human level accuracy (88%) and built large Multi-NER models. Submitted Industry ready - Dockerized System Demonstration in AMIA, 22 to be conducted in Washington, DC.

Tabular NLI Knowledge Addition and Inference. Developed custom RoBERTa and BiLSTMs to integrate external KGs like Concept-Net and WordNet. Improved benchmark performance by 7.5% on Infotabs, a tabular NLI dataset and published TKBLSTM paper. Worked on development of contextual embeddings for better sentence similarity and premise re-ranking.