Manav Kaushik Email & Phone Number

India

Hyderabad, Telangana, India

- Worked on Speaker Characteristics Estimation (such as height, age, accent, etc.) using speech analysis to aid professional forensics & human-computer interaction.- Prepared Attention/ Transformer-based deep learning algorithms for weighing different phonemes to enhance estimation performance and understand the predictive power of individual phonemes.- Worked to extract and train embeddings using Triplet & Center Loss to further the discriminative capabilities of the proposed algorithms. - Used self-supervised learning algorithms such as Wav2Vec2 for better feature extractions.

Pilani, Rajasthan, India

Developed a Deep Learning models (using Convolutional Neural Network & Transfer Learning) to detect signs of drowsiness in car drivers. 'Drivers Drowsiness Detection Dataset' (by NTHU Computer Vision Lab) was used and fed to the model after augmentation & pre-processing. The fnal fne-tuned model (with VGG16 weights pre-trained on ImageNet Dataset), achieved a highly competitive accuracy of over 72%.

Bengaluru Area, India

- Interned with the Big Data & Advanced Analytics Division of American Express.- Developed an Automated Machine Learning framework (named One AI) for automated data extraction, dynamic visualizations, ML modelling, and ready to use prediction pipeline using Python & HiveQL.- 'One AI' incorporates over 25 most popular ML & statistical algorithms such as Logistic Regression, Gradient Boosting, Random Forest, XGBoost, ANNs, and many more.- Successfully tested 'One AI' on one of Amex's use cases enhancing the error detection rate from 1% to 30% translating into the reduction of man hours from over 2000 to less than 70 hours per month.

Delhi Area, India

Developed an Early Warning System for Financial Crisis in emerging economies using different techniques of predictive modeling: Multinomial Logit Modeling & Contemporary Machine Learning Algorithms. The model combines both Macroeconomic as well as Stock Market Variables to achieve higher accuracy. Historical data for several emerging economies, collected fromThomson Reuters & Bloomberg, are used in the model. The model can successfully predict Financial Crisis for 10 major Emerging Economies of the world 2 quarters beforehand with an accuracy of over 90%.

Ahmedabad, Gujarat, India

Implemented a Deep Learning-based model for Short-Term Electricity Load Forecastingaddressing the challenge of improving the accuracy in electricity distribution utilityin presence of volatility and mobile loads. The developed model demonstrated an accuracy over 96%, a considerable improvement of more than 5% against current levels. The project involved collecting real world data from the web (consisting of over 100K data points) & leveraged the concepts of Data Science: Machine Learning (Artificial Neural Networks), Programming, Data Handling & Visualization.