Websocket-based client-server application for Anomaly detection on sensory data
Delivering an end-to-end client-server websocket-based application which can run anomaly detection algorithms for sensory data collected by wearable sensors to alert user.
Fetching of data using websocket application, real-time streaming to the server to run the model and reporting to the user in web-application format.
May 2020 - Present
Graduate Student Research
Microsoft
Interpretable Tree models for Anomaly Reasoning
Collaborated with Microsoft’s AI Development Acceleration Program for a comparative study between the open source SHAP solution for model-agnostic interpretations and Griffon’s tree-based reasoning solution created by Microsoft
If possible, extend SHAP to be used to interpret Griffon's reasoning by replacing Tree-based interpreter package native to Griffon to improve performance.
January 2020 - Present
Senior Software Engineer
Samsung Research Institute Bangalore
Pattern Analysis for Device Usage Statistics
Mined frequent device usage patterns using Apache Spark Framework for users from the SmartThings data on AWS EMR instances. Sharded Data on MongoDB to enable efficient GDPR implementation.
Implemented end-to-end Scala application running on Spark framework by using association rule mining. Project commercialized in 2019 with the release of Samsung Galaxy Note 10.
Lightweight User Presence Detection backend on powered Embedded devices
Designed Neural Network based Voice Activity Detection in C++ on TizenOS to detect human presence at Home for Smart Speakers. Used MRCG features and Tensorflow Lite to optimize time and memory. Conferred with performance award for reducing inference time by 5-folds.
Behavioral AI framework to enable user personalization in Social Robots
Designed and Implemented Behavioral Intelligence framework in C++/Python by jointly employing Neural Network alongside Q- Learning for implementation of User Personalization among robots to achieve 3X faster convergence with twice the accuracy against standard Reinforcement Learning Techniques.
July 2017 - August 2019
Skills
Programming Languages & Tools
C/C++
Python
Java/Javascript
Scala
R
MATLAB
Data & Machine Learning Tools
Apache Spark
PySpark
Hadoop
Numpy
Tensorflow
AWS
SkLearn
Keras
Miscellaneous
Agile
Git
LaTeX
MySQL
Projects
OSMI Mental Health Data Exploratory Data Analysis
Exploratory Data Analysis on the OSMI Mental Health Data to understand how mental health is viewed in the Tech industry, the spread of mental disorders among the employees and to gauge the system present to tackle these conditions.
D3 based visualizations based on the survey results of 3 different years implemented to gain a multi-faceted view of our data. Site hosted at shezanmirzan.github.io/DataVis-Mental-Health
Deep Multiple Instance Leaning based Video Classification
Developed Anomaly detection algorithm for classifying real - Surveillance videos that spanned across different scenes.
Converted the classification problem to a regression task by extracting C3D features and feeding it to deep Multiple Instance
Learning based architecture to get higher scores on video segments that contained anomaly.
Tried different model architectures and feature extraction and compared ROC curves to decide on the best model.
Tracking of Multiple Skin-Colored Objects Under Occlusion
Developed Real-time tracking of skin coloured objects framework in MATLAB using object hypothesis tracking. removal and
synthesis.
Found applications in the field of hand-gesture recognition to detect gestures involving occlusion
Patents & Publications
Evaluating Tree Explanation Methods for Anomaly Reasoning: A Case Study of SHAP TreeExplainer and TreeInterpreter