Experience

W365 Hibernate Project:

Developed a combination of deep learning transformer-based approaches and statistical models in collaboration with the Microsoft Research Substrate team for the W365 Hibernate Project, resulting in a 40% reduction in COGS per user and annual savings of $1M
with 90% efficiency.
Additionally, I streamlined MLOps processes, resulting in modularized code and efficient model deployment using Azure ML, Synapse and Kusto(SQL) and effectively communicated insights from extensive research on model and user behavior to stakeholders.

Frontline Prestart Project:

Successfully led the end-to-end machine learning model implementation of predicting user login times, ensuring users could log in within 60 seconds of starting the machine. This solution was effective for an estimated 60% of Frontline users with a minimal 12% COGS loss.
Additionally, I developed the required metrics and an efficient MLOps solution in Azure ML for faster model iteration and created the first version of user segmentation for rollout and activity pattern analysis.

Copilot Scenarios:

Implemented a working solution to enhance the efficiency of existing data processing workflows by utilizing large language models (LLMs) by working on fine tuning the LLM and working extensively on Prompt Engineering.
Leveraged GPT 3.5 to accomplish incident ticket summarization and topic model labeler with also including validation of LLM generated outputs.

Led implementation of data science workflows from data extraction, statistical inference, machine learning models, dashboards to model deployment. A senior lead to the company’s machine learning product Cortex.

Managed and mentored data science interns and process engineers. Played a pivotal role in growing the data science team at the company by engaging with Product owners, Design and Data Engineers.
Collaborated with key stakeholders to influence the product roadmap and successfully got buy-in from prospective clients for machine learning proof of concepts (PoC) on Cortex (machine learning platform).

Abnormal behavior detection and Recommendation diagnostics engine:

Full stack lead on developing the entire data science framework with 10+ different burner heaters including data ingestion, exploratory analysis , hypothesis testing , model building, model deployment and model maintenance.
Model building using Gaussian mixture models, Autoencoders and Copula based outlier detection.
Model deployment using SageMaker studio , Apache Airflow and Causal inference for higher anomaly points.
This is currently one of the highest revenue generating products in the company, with each anomaly detected successfully reducing the operating cost by $50K each time and reducing maintenance turnaround time by 6 months.

Vegetation Failure forecasting and inference:

Predictive analytics by forecasting external failures in multiple electrical plants saving more than $10M in operating costs and possible wildfires since deployment. Utilizing raw signal processing and crucial feature engineering built multiple ML classification models with custom metrics for loss function with Probabilistic Neural Networks for key inference.
Data Engineering pipeline for weather data, imagery and tabular data were implemented using AWS Sagemaker and EC2 instances and built SQL queries for efficient data quality and model performance.

Machine learning product (Cortex):

Developed several machine learning algorithms from scratch for the machine learning product Cortex including KNN, Random Forests, Neural Networks, Bayesian Networks, LSTM, Model Selection.
Implemented and modified Deep Auto-Encoding Gaussian Mixture Models from scratch using Python and Tensorflow. Scaled and deployable the model to tackle an unsupervised learning problem for anomaly detection. This PoC made it into the Cortex product to be used by non-data science users.
Architected the pipeline for automation of data ETL and data preparation workflows. Established a monitoring framework to effectively visualize model outputs on being deployed.
Developed key customer and product metrics that helped scope machine learning projects and ensured better product experience for the end user using A/B testing and statistical modeling.

Type of Failure Detection (Energy grid):

Built a robust multi-classification model to detect the type of failure, enabling maintenance crews to schedule operations in a timely manner effectively reducing unit down times.
A 1-dimensional CNN for waveforms and Xgboost on tabular data were combined to give a more balanced accuracy score.
Additional clustering with DBSCAN and Feature Engineering with AutoEncoders were used to effectively improve both the model and the data exploration.

Smart Combustion (Solex):

Real time forecasting, control and recommendation of NOx (Nitric oxides) values for the first ever single NOx burner called Solex.
Lead Data Scientist in developing the solution from data extraction to model development and scaling using SVR, Xgboost
The control optimization was executed using Markov Decision Process and Model Predictive Control.
The model was deployed widely on various burners using Programmable Logic Controller edge computing and saved close to $1M in costs for the end client in the initial year of adoption.

Unsupervised event prediction (Invista):

Researched and implemented an unsupervised approach to detect and predict system failure events.
The challenges were met by implementing a combination of deep feature engineering and custom loss functions on LSTM networks and clustering some known events using DBSCAN and Deep AutoEncoders with Gaussian Mixture Models.

Developed Natural Language Processing based topic modeling for document classification using Latent Dirichlet allocation (LDA).
Implemented a custom named entity recognition to recognize different entities in a document for a cybersecurity client.

Implemented various machine learning methods, feature engineering techniques to find patterns in the clinical data using Python.
Developed a model to select the top most suitable medicines based on chemical composition with very low negative side effects.

Assisted with statistical analysis and inference of clinical trial data to assess the impact of drugs on alleviating anxiety disorders.

Worked as part of the research team of 3 members, to analyze the performance of sensors to be deployed in an IoT network.
Implemented a clustering algorithm to group individual sensors based on parameters such as temperature, voltage output, effect of physical parameters on sensors etc. by modeling real-time time series sensor data using R and SQL
Designed an anomaly detection model that helped identify outliers which represented faulty sensors using Density based detection.