Projects

View on LinkedIn →

2023 CMU

Intelligent Traffic Monitoring

Jan 2023 – Dec 2023

Working in collaboration with Bosch on improving anomaly detection in traffic videos. Integrating Text-Conditioned Object Detection and Automatic Segmentation in order to detect objects in the video frame. Detected objects to be fed to the Knowledge graph in order to build an ontology that can be utilized for the prediction pipeline.

Knowledge Graphs Computer Vision Deep Learning
2023 CMU

Multimodal Emotion Detection

Jan 2023 – May 2023

Utilized M2FNet multimodal architecture in order to perform emotion detection in videos. Improved emotion detection accuracy by 5% by integrating humor features in a multimodal pipeline involving audio, text, and video. Enhanced the feature extraction process for the various modalities in order to improve the multi-head attention-based model. Provided an in-depth comparative analysis of the variation in performance when different modalities are used to make predictions.

Feature Extraction Feature Engineering Multimodal ML
2023 CMU

Sensor Fusion and Tracking for Autonomous Racing

Mar 2023 – May 2023

Developed a forecasting model in order to track and forecast the future position of a car utilizing data fused from point clouds and 2D images. Ensured that the developed model could be run with limited resources and on the edge to provide real-time predictions. The developed model was based on the LSTM architecture with a Time Distributed Layer that could predict k timesteps based on the previous n timesteps. Achieved accuracies of 93% and 63% for 2D and 3D bounding boxes respectively.

Feature Extraction Multimodal ML Sensor Fusion
2022 CMU

End-To-End Cloud Microservice

Aug 2022 – Dec 2022

Developed an end-to-end user recommendation service analyzing Twitter users and hosted the microservice on AWS. The service involved a backend SQL database which was populated with big data analysis using Spark. Kubernetes was used to host the cluster and Helm was used to install the developed docker images as pods. The service achieved an RPS in excess of the recommended 10,000 RPS when tested under load. Additionally, developed microservices in order to encode and decode QR codes and achieve a performance of 65,000 requests per second.

Docker Kubernetes Helm AWS Spark
2021 PES University

Automatic Question Answer Generator

Jan 2021 – Dec 2021

Developed an automatic question answer generation pipeline using Transformer Networks. To improve span extraction, sentences were encoded using the Universal Sentence Encoder and similarity scores were calculated to implement a coherence matrix through which the sentences were grouped and abstractively summarized. Additionally knowledge graphs were utilized to enhance Domain Knowledge.

Publication →
Feature Extraction Machine Learning Knowledge Graphs
2020 PES University

Source Code Plagiarism Detection

May 2020 – Dec 2020

A machine learning based source code plagiarism detector used to detect plagiarism in C programming assignments. Various approaches are experimented with including Milepost GCC for feature extraction followed by training machine learning models, training a Recurrent Neural Network with LSTM units on part of the Linux Kernel to extract features, and using the SOTA Transformer Network for feature extraction.

Publication →
Data Collection Feature Engineering Machine Learning