Sivaram GS | Senior Data Science Engineer

Hi, I'm Sivaram GS.

A

Self-driven, quick starter, passionate Data Science Engineer with a curious mind who enjoys solving a complex and challenging real-world problems.

About

As a Senior Data Science Engineer with over 8 years of experience, I apply my passion and expertise in data science and generative AI to create innovative and impactful solutions for the media and entertainment industry. I also have extensive experience in data engineering and optimizing data pipelines, delivering analytics solutions for major sports leagues, and deploying inSITE (Evertz's big data analytic solution) for big broadcasters.

Languages: Python, Pyspark.
Databases: MySQL, PostgreSQL, Pinecone, PG_Vector, Elasticsearch, Redis.
Platforms: AWS, Azure, Databricks, Snowflake, Spark, Hadoop, Kafka.
Libraries: NumPy, Pandas, Scikit-learn, NLTK, Matplotlib, Plotly, MLflow, Opencv.
Frameworks: Keras, TensorFlow, PyTorch, Langchain, Chainlit, Streamlit, Logstash.
Tools: PowerBI, Git, Docker, Jira, Confluence, Notion.
Technologies: Machine Learning, Deep Learning, NLP, LLM, GenAI, MLOps, LLMOps.

Experience

Evertz India Pvt Ltd

Senior Data Science Engineer

Spear-headed the team that built and launched evGPT - a chatGPT for Evertz powered by Large Language Models (GPT-4 Turbo) using Azure OpenAI. Collaborated with different product owners to ingest quality data and fine-tuned the LLM resulting in 70% accurate response.
Delivered analytics solution for the major sports leagues: NBA, MLB, NHL on the first party data collected by Evertz Easelive platform by applying machine learning techniques and achieved 50% increase in users viewing duration, 68% increase in poll response rate and 56% increase in user engagement.
Built optimized data pipelines to read large volume of real-time Easelive data and stored in customer destinations: snowflake, databricks reducing reports creation time by 80%.
Designed, configured, and managed inSITE (Evertz Bigdata Analytic Solution) for customers: DU Samacom (UAE), Fox (US), Rogers (Canada)

April 2020 - Present | Bengaluru, India

Dell Technologies

Software Engineer 1

Performed data cleaning and built analytics for DELL’s Digital Lifecare that powers the Govt of India’s Ayushman Bharat scheme for preventive screening of diabetes, hypertension, and cancers detecting fraudulent signatures in people data collected by Primary Health Centers workers.
Containerized DELL EMC Solutions Enabler Product by reducing the OS image size by 70%.
US Patent-11256553 granted under the class Computer Network Monitoring – “INTELLIGENT SYSTEM TO PRIORITIZE SERVERS FOR ENVISAGED REQUESTS”.

June 2017 - March 2020 | Bengaluru, India

Tata Elxsi

Engineer

Delivered a Video on Demand workflow for Comcast by optimizing RTVOD, C3 and C7 windows with full ad insertion support reducing quality check and content packaging time by 60%.
Configured and Deployed File based Media Asset Management Playout systems for Linear and Non-Linear workflows.

July 2016 - June 2017 | Bengaluru, India

Projects

                
                Inference of LaMini with Amazon SageMaker using Hugging Face DLC
                  
ProjectTools:LaMini-T5-738M, Amazon Sagemaker, Hugging Face Deep Learning Container, Python 3.10 , Pytorch, Gradio
Created a domain in Amazon Sagemaker for managing and organizing the machine learning workloads.
Created a notebook instance (ml.c5.2xlarge) which is pre-installed with popular libraries for machine learning and to keep the compute environment secure for LaMini-T5-738M.
Installed the required libraies and deployed the model with instance_type="ml.g4dn.xlarge" using Hugging Face Deep Learning Container.
Noted the Endpoint created under the inference section in Amazon sagemaker.
Tuned the model with hyperparameter payload and integrated the sagemaker endpoint with a gradio.
Launched the gradio app with a public URL for LaMini-T5-738M text2text-generation task inference.

                Serverless Blog Generation with Titan Model Using Bedrock in Lambda
                  
ProjectTools:Postman, Amazon API Gateway, AWS Lambda, Titan Text G1 Express Model , Amazon Bedrock, S3 Bucket, Cloudwatch
Requested and got the access granted for Titan Text G1 Express Model in Amazon Bedrock.
Noted model id and API request for Titan Text G1 Express in Amazon Bedrock.
Created a API Gateway to trigger AWS lambda function and added route to make a POST API request.
Written the python code to invoke Titan Text G1 Express model for blog generation of user topic.
Made a POST API request from POSTMAN to Amazon API gateway endpoint to trigger the AWS lambda.
Monitored the cloudwatch logs and saved the generated blog as a text file in S3 bucket.

                Custom WebSearch LLM Agent with Llama3 Instruct 8B Using Ollama
                  
ProjectTools: Llama3 Instruct 8B llm, Ollama, Serper Google Search API, Python 3.10
Pulled llama3 instruct 8B fine-tuned LLM in ollama (framework to run llm models locally) 
Used serper as the primary tool for the search engine query intened for google search.
Created planning agent to take user custom query as input and generate searchable queries.
Created Integration agent to validate planning agent output and make a google search.
Integration agent visited the sites and scraped the best pages to compile a response.
Made a quality assessment on the response compiled and generated the final response if criteria met.

                Sentiment Analysis on Bing News Data Using Microsft Fabric
               
ProjectTools: Azure Bing Search V7 API, Fabric Data Factory, Synapse Data Engineering, Synapse Analytics, PowerBi, Data Activator, Reflex.
Ingested data from Bing V7 API using Data Factory copydata activity.
Transformated data to a curated Delta Table with incremental loading.
Done Sentiment Analysis on the news and classified it as Positive / Negative / Neutral/ Mixed.
Used DAX queries in PowerBi to create new measures and configured alerts using Data Activator Relex.
Tested the complete flow and reviewed pipeline run results in Monitoring Hub.

                Stream Processing in Databricks Lakehouse with Unity Catalog
              
ProjectTools: Azure ADLS Gen2, Databricks, Unity Catalog, Azure Devops, Key Vault.
Created storage containers in ADLS Gen2 to store metadata, managed and unmanaged tables.
Supported Batch and Streaming Worklows to ingest data.
Designed and Implemented a secure Lakehouse Platform with Unity Catalog.
Generated Workout BPM and Gym summary Datasets for analytics.
Automated deployment pipeline in Azure Devops.

                End to End Azure Data Engineering Real-Time Project
             
ProjectTools: Azure Data Factory, ADLS Gen2,  Key Vault, Databricks, Synapse Analytics, PowerBi, Active Directory.
Started with data ingestion from on premises SSMS, into the Azure environment. Utilized Azure Data Factory for orchestrating and automating this process.
Used Azure databricks to transform the RAW data.
Used Azure Synapse Analytics to load the clean data (gold tables). 
Used PowerBi Analytics to create an interactive dashboard.
Integrated Azure Active Directory with key-vault for user authentication and authorization.

                  Generative Adversial Network for Fashion Mnist with Tensorflow
                 
ProjectTools: Fashion Mnist Dataset 3.0.1, Tensorflow 2.16
Preprocessed the dataset by scaling images through caching, shuffling, batching, and prefetching. 
Built the Generator that transforms random noise into realistic fashion images.
Built the Discriminator that distinguishes generated fake fashion images from real images.
Trained the GAN Model for 100 Epoch through backpropagation on Generator and Discriminator.
Reviewed the Model Performance by plotting discriminator and generator losses for 100 epochs.

patent

Intelligent system to prioritize servers for envisaged requests

Patent number: 11256553
Date of Patent: Feb 22, 2022

A workload manager uses on-band and off-band metrics to select a host server in a cluster to handle a connection request. The on-band metrics include CPU usage, memory usage, and vulnerability metrics. The off-band metrics include hardware component error logs. Utilization and vulnerability scores are calculated for each host server from the on-band metrics. A reliability score is calculated for each host server from the off-band metrics. A health score for each host server is calculated from the vulnerability and reliability scores. The health score is used to exclude unhealthy host servers from consideration. A priority score is calculated for each host server from the utilization, vulnerability, and reliability scores. The host server that has not been excluded and has the greatest priority score is selected to handle the connection request.