Successfully created end to end automated speech recognition engine called music-on-demand using wav2letter for African accent. o Achieved word error rate(WER) less than 10%.

Developed a Document Search Engine using Deep Learning and Elastic search, which quickly searches a context based response for a query, from a large number of documents. The system highlights top 'k' relevant responses and navigates to their actual positions in the document

Successfully built end to end goal-oriented botfor one of Australia’s largest energy companies to route customer calls to correct LOB (electricity, gas) as per user’s response and auto scaled it using google cloud and google Kubernetes engine.

o Bot is capable of closing 50-60 Daily call on its own without intervention of live agents and transfer rest of the calls to agents if it fails to close the call at some node.

This is a research oriented project specially undertaken for Vodafone. The components include -
- Entities Extraction: This was done using BERT Bi-LSTMAttnCRF and the model was tested using maluba dataset by
Microsoft.
- User action and Intent prediction: Built custom classifier model on top of pretrained BERT embeddings
- State tracker: Creation of vector based state is done which handles the current position of conversation
along with entire conversation history.
- Policy: Used Deep Reinforcement Learning for building DQN which handles the action that needs to be
performed. Currently working on many more policy algorithms to improve user experience.
- Response selection: using of Facebook’s cross encoders what made this part outstanding from conventional NLG’s.

The main objective of this project was to automate car insurance which is a viable process since it involves lots ofsteps and intermediate parties to successfully process a single claim. Therefore, we aim to design a system thatautomates the processing repair claims by employing different deep learning techniques, so as to alleviate thedependence on manual inspection and the bias invariably introduced by a human surveyor.
- Build deep CNN models for instance segmentation and localization of various parts along with classification of type of cars.
- Deep learning models include Mask R-CNN, PANet, VGG-16. Transfer learning and ensembling of models has been done to achieve better results with less amount of data.