1. Text to SQL Translation:

The main objective was to enhance the user experience within Algo chatbot by incorporating natural language to SQL generator using state-of-the-art language models such as NL2SQL-BERT & Seq2SQL and achieved 84% accuracy. Currently, testing it for multiple retailers of Microsoft such as Best Buy.

2. Context in Group Chat:

To achieve considerable accuracy on long contextual conversations within group chat, I worked on multiple methodologies in NLP and deep learning including multi-intent dialogs, hierarchical attention, named entity recognition and contextual word embedding.

1. Context-aware Anomaly Detection:

The main objective was to enhance the real time predictions of recommender system by detecting and eliminating the effect of contextual and behavioral anomalies. Novel method was developed using machine learning and big data analytics and achieved 8% overall performance increase.

2. Clustering:

To improve the existing clusters for customer-calling groups, domain-aware representations were learnt on the data set of AT&T US and T-Mobile . To achieve the generalization in the approach, few shot learning techniques were used for other similar clients.

Final year research thesis under Dr. Irfan Younas & Dr. Mubasher Baig - This project aimed to solve the problem of contextual understanding in the chat bot conversation to carry out near-human negotiations. The main intent was to solve the complexity of contextual negotiations on e-commerce and freelancing websites. It was achieved using Bayesian machine learning, NLP and reinforcement learning techniques including Seq2Seq, BERT models.

• Road Scene Classification for Self-driving Cars using Spatio-temporal techniques
• Incorporating the context of road scenes of both developing and developed countries in a classification model for self-driving cars was the major goal of this research.
• Collected the Road scenes data comprising markets, bridges, highways, and intersections.
• Spatial classification using transfer learning approaches (Inception-ResNet-v2 and ResNet-50 models).
• Handled noise and scene transitions using Spatio-temporal techniques, resulting in accuracies up to 91%.