San Francisco Bay Area
2K followers 500+ connections

Join to view profile

About

Software Engineer specializing in Generative AI/Machine Learning, Backend engineering…

Activity

Join now to see all activity

Experience & Education

  • Uber

View Pranav Kumar’s full experience

See their title, tenure and more.

or

By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.

Licenses & Certifications

Volunteer Experience

  • Board Member

    Leo Club of NITT Rocktown

    - 1 year 8 months

    Social Services

    LEO Club of NITT Rocktown, a social club of NIT Trichy, is one of the most famous and established clubs in the campus. The LEO club of NITT Rocktown is affiliated to the LIONS club international.
    The objective of the club is to promote social and service activities among the Nittians which will develop the individual qualities of Leadership, Experience, and Opportunity and to unite its members in friendship, fellowship, and mutual understanding.

  • Member

    Aarushi - The Social Responsibility Group

    - 1 year 6 months

    Social Services

    Was a part of the fundraising division, involved in the marketing of the organization.
    Primarily contributed to the construction of old age homes in and around Trichy.

Publications

  • Applications of Sequence to Sequence Models for Technical Support Automation

    2018 IEEE International Conference on Big Data

    Juniper Networks, Inc. offers hardware products and software services to its enterprise customers. Due to the nature of it’s business, Juniper Networks, Inc. is deeply invested in providing the best customer support and as part of the support automation team, our goal is to optimize the company’s efforts towards it. For this purpose, alongside other initiatives, we leverage deep learning based sequence to sequence models wherever we see fit. In this paper, we discuss two such models: a…

    Juniper Networks, Inc. offers hardware products and software services to its enterprise customers. Due to the nature of it’s business, Juniper Networks, Inc. is deeply invested in providing the best customer support and as part of the support automation team, our goal is to optimize the company’s efforts towards it. For this purpose, alongside other initiatives, we leverage deep learning based sequence to sequence models wherever we see fit. In this paper, we discuss two such models: a conversational chatbot to help answer some technical questions for our customers, and a text summarizer to condense the large text in our support tickets and other articles. These two models are designed using bi-directional recurrent neural network (Bi-RNN) architectures for content understanding and were customized to fit the domain-specific nature of our data. First, we discuss our efforts towards data preparation. Then, we explain our model design, customization and evaluation mechanisms. Finally, we provide the preliminary results and share the potential impact our models will have on our business. Our initial results have BLEU score of 0.21 for text summarizer which is 16% better than our baseline model. Our chatbot passed the eye-tests of our subject matter experts.

    Other authors
    See publication

  • A Real-time Robust Facial Expression Recognition System using HOG Features

    IEEE International Conference on Computing, Analytics and Security Trends, Pune

    This paper presents a facial expression recognition
    framework which infers the emotional states in real-time, thereby
    enabling the computers to interact more intelligently with people.
    The proposed method determines the face as well as the facial
    landmark points, extracts discriminating features from suitable
    facial regions, and classifies the expressions in real-time from
    live webcam feed. The speed of the system is improved by
    the appropriate combination of the detection…

    This paper presents a facial expression recognition
    framework which infers the emotional states in real-time, thereby
    enabling the computers to interact more intelligently with people.
    The proposed method determines the face as well as the facial
    landmark points, extracts discriminating features from suitable
    facial regions, and classifies the expressions in real-time from
    live webcam feed. The speed of the system is improved by
    the appropriate combination of the detection and tracking
    algorithms. Further, instead of the whole face, histogram of
    oriented gradients (HOG) features are extracted from the active
    facial patches which makes the system robust against the scale
    and pose variations. The feature vectors are further fed to a
    support vector machine (SVM) classifier to classify into neutral or
    six universal expressions. Experimental results show an accuracy
    of 95% with 5 folds cross-validation in extended Cohn-Kanade
    (CK+) dataset.

    Other authors
    See publication

  • An Unsupervised Approach for Overlapping Cervical Cell Cytoplasm Segmentation

    IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES) -2016 , KL, Malaysia

    The poor contrast and the overlapping of cervical
    cell cytoplasm are the major issues in the accurate segmentation
    of cervical cell cytoplasm. This paper presents an automated
    unsupervised cytoplasm segmentation approach which can effectively
    find the cytoplasm boundaries in overlapping cells.
    The proposed approach first segments the cell clumps from the
    cervical smear image and detects the nuclei in each cell clump.
    A modified Otsu method with prior class probability is…

    The poor contrast and the overlapping of cervical
    cell cytoplasm are the major issues in the accurate segmentation
    of cervical cell cytoplasm. This paper presents an automated
    unsupervised cytoplasm segmentation approach which can effectively
    find the cytoplasm boundaries in overlapping cells.
    The proposed approach first segments the cell clumps from the
    cervical smear image and detects the nuclei in each cell clump.
    A modified Otsu method with prior class probability is proposed
    for accurate segmentation of nuclei from the cell clumps. Using
    distance regularized level set evolution, the contour around each
    nucleus is evolved until it reaches the cytoplasm boundaries.
    Promising results were obtained by experimenting on ISBI 2015
    challenge dataset.

    Other authors
    See publication

Projects

  • A Versatile Online System for Person-specific Facial Expression Recognition

    This project represents a framework for online recognition of facial expressions. The workflow
    includes the detection of face and facial landmark points, extraction of discriminating features
    from suitable facial regions, and classification of expressions – all in real-time from live webcam
    feed. The key to the improvement of the real-time performance is the appropriate combination
    of the detection and tracking algorithms. Further, instead of using features of the whole face…

    This project represents a framework for online recognition of facial expressions. The workflow
    includes the detection of face and facial landmark points, extraction of discriminating features
    from suitable facial regions, and classification of expressions – all in real-time from live webcam
    feed. The key to the improvement of the real-time performance is the appropriate combination
    of the detection and tracking algorithms. Further, instead of using features of the whole face, it
    uses the histogram of oriented gradient (HOG) features extracted from the active facial patches to
    reduce the computational complexity. This also makes the system robust against the scale and pose variations. The resting face of some people may look angry or happy or sad. Therefore, the relative change of facial landmark positions as well as the HOG features from the neutral face features, are used to classify the face into neutral or six universal expressions. The framework adapts itself to new user faces by extracting the neutral features of the user during the time of execution and using it to train the classifiers. Promising results were obtained on publicly available databases which proves the effectiveness of the proposed approach.

    Other creators
    See project

  • Multi-Object tracking on NVIDIA Jetson TX2 in Subterranea

    -

    This project is mainly about developing a high-performance real-time object detection framework in an embedded AI computing device known as Nvidia Jetson TX2.

    See project

  • Movie Recommender system using Spark and ElasticSearch

    -

    In this project, we propose an approach to provide real-time recommendations, by constructing a large-scale recommender engine using big-data technologies like Apache Spark and ElasticSearch. This approach is highly scalable as it involves frameworks based on distributed systems and cluster computing. This project is very useful for moviegoers to choose the movie that satisfies their preferences and so the motivation is of very high utility. The output of this project gives the user a list of…

    In this project, we propose an approach to provide real-time recommendations, by constructing a large-scale recommender engine using big-data technologies like Apache Spark and ElasticSearch. This approach is highly scalable as it involves frameworks based on distributed systems and cluster computing. This project is very useful for moviegoers to choose the movie that satisfies their preferences and so the motivation is of very high utility. The output of this project gives the user a list of movies that he/she will want to see. Say we want to see movies similar to “Men in Black”. The system will recommend sequels of Men in Black such as “Men in Black 2” or even similar action genre movies like “Spider-Man”.

    Other creators
    See project

  • Copter Bot based on Deep Q-Learning

    -

    The objective is to develop a Deep Neural network model to automate the gameplay in
    Copter game. The Copter is a classic side-scrolling game where the agent must successfully
    navigate through a cavern.
    • A reinforcement learning algorithm called Q-Learning is utilized as it has
    been proven that for any finite Markov Decision Process, at a given state
    Q-learning can find an optimal policy such that the reward is maximum
    by taking into account the successive states.
    • Our…

    The objective is to develop a Deep Neural network model to automate the gameplay in
    Copter game. The Copter is a classic side-scrolling game where the agent must successfully
    navigate through a cavern.
    • A reinforcement learning algorithm called Q-Learning is utilized as it has
    been proven that for any finite Markov Decision Process, at a given state
    Q-learning can find an optimal policy such that the reward is maximum
    by taking into account the successive states.
    • Our goal is to progress from a simple Multi-layered perceptron (MLP) to
    experimenting with a convolutional neural network with a couple of fully
    connected layers and more complex additions.
    • We also plan to implement Experience Replay which will allow our
    network to train itself using stored memories from it’s experience.

    Other creators
    See project

  • MBTI Type prediction and analysis

    -

    To develop an efficient system that predicts MBTI (Myers Briggs Type Indicator) personality types of persons based on their writings/posts. According to MBTI personality type system everyone is classified into one of sixteen distinct personality types across four axis,
    Introversion(I) – Extroversion(E)
    Intuition(N) – Sensing(S)
    Thinking(T) – Feeling(F)
    Judging(J) – Perceiving(P)
    (Examples types look like INTP, ENTJ, ISTJ, ESFP...)
    And also find and discover interesting…

    To develop an efficient system that predicts MBTI (Myers Briggs Type Indicator) personality types of persons based on their writings/posts. According to MBTI personality type system everyone is classified into one of sixteen distinct personality types across four axis,
    Introversion(I) – Extroversion(E)
    Intuition(N) – Sensing(S)
    Thinking(T) – Feeling(F)
    Judging(J) – Perceiving(P)
    (Examples types look like INTP, ENTJ, ISTJ, ESFP...)
    And also find and discover interesting insights/patterns from the available dataset.

    See project

  • Twitter Images auto-trending-hashtagger

    -

    An online hashtag-suggesting web application that runs on Kafka, Storm, and Cassandra which predicts relevant hashtags that are trending in the current week and day in Twitter with respect to the images that uploaded.

    Other creators
    See project

  • Next Basket Prediction

    -

    The Dataset that was used is from Kaggle hosted competition ' Instacart Market Basket Analysis'.
    Implemented a less memory intensive solution using LightGBM model with multiple statistical features learned from the given data, on each type of entities like user, product, order. Developed an internal structure of features like thresholds for re-order probabilities and some more features for a final Gradient boosting classifier with an intention to maximize the F1 score.

    See project

  • Recommender Systems For Cold-Start Items : A Deep-Learning And Collaborative Filtering Based Approach

    -

    This thesis implements a combination of several methods to overcome the cold-start items problem by leveraging the content features of the items under consideration using Deep Learning Networks and Content-Based Filtering. A Stacked Denoising Auto-Encoder (SDAE) deep network is trained to generate feature representations from content-based item vectors and subsequently to estimate the ratings of Cold Start Items. This is then supplied to the Singular Valued Decomposition++ (SVD++) Collaborative…

    This thesis implements a combination of several methods to overcome the cold-start items problem by leveraging the content features of the items under consideration using Deep Learning Networks and Content-Based Filtering. A Stacked Denoising Auto-Encoder (SDAE) deep network is trained to generate feature representations from content-based item vectors and subsequently to estimate the ratings of Cold Start Items. This is then supplied to the Singular Valued Decomposition++ (SVD++) Collaborative Filtering model for Rating Prediction. Experiments are performed on the MovieLens dataset, comparing movie content information between items to estimate ratings of movie that do not have any ratings yet.

    Other creators
    See project

Honors & Awards

  • Young Scientist Award

    IECBES-2016

    Best Paper in Biomedical Signal and Image Processing,
    IEEE EMBS CONFERENCE ON BIOMEDICAL ENGINEERING AND SCIENCES 2016. KL, Malaysia

Languages

  • English

    Full professional proficiency

  • Tamil

    Native or bilingual proficiency

  • Telugu

    Limited working proficiency

  • Hindi

    Limited working proficiency

  • Malayalam

    Elementary proficiency

Organizations

  • Vortex

    Publicity Manager

    -

  • MIRSAA

    Alumni Member

    -

    Conducting coding events for high school students from Maharishi International Residential School, Sriperumpudhur, Kancheepuram, Tamilnadu

More activity by Pranav Kumar

View Pranav Kumar’s full profile

  • See who you know in common
  • Get introduced
  • Contact Pranav Kumar directly
Join to view full profile

Other similar profiles

Explore more posts

Explore top content on LinkedIn

Find curated posts and insights for relevant topics all in one place.

View top content

Add new skills with these courses

See all courses