PROJECT PRESENTATION

Esha Poddar

Personal Projects

start

Index

Future Goals

About Me

Thesis

Work Projects

About Me

Esha Poddar

Data Professional | Tech Enthusiast LinkedIn: Esha Poddar Github: eshapoddar Email: eshapoddar@gmail.com | Phone: +44 07768282853

Skills

• Versatile Technologist: Proficient in Python, SQL, Terraform, Docker, Kubernetes.
• Cloud & DevOps: Experienced with Google Cloud, AWS, Jenkins, Airflow, Terraform.
• Data Visualization Expert: Google Data Studio, Tableau, Power BI.
• Version Control: Git & Git Workflows.

01

Work Projects

Product/Services

Infrastructure as Code for Research Environments:

CICD ETL Pipeline for Retail Chain:

• Engineered a Continuous Integration/ Continuous Deployment (CICD) ETL pipeline.
• Designed specifically for a retail chain, leveraging the power of BigQuery.
• Streamlines data extraction, transformation, and loading processes for retail analytics and insights.

• Developed an Infrastructure as Code (IaC) platform catering to researchers' needs.
• Enables seamless setup of cloud environments tailored to specific research topics.
• Automates infrastructure setup, including VM provisioning and Jupyter Notebook configuration based on chosen datasets.

02

Masters Thesis

IMPLEMENTATION OF MLOPS USING NO-CODE AI PLATFORMS

Situation

Democratizing AI: Integrating No-Code Platforms with MLOps

Preface of the thesis

MLOps, an amalgamation of Machine Learning (ML) and Operations (Ops), streamlines the development and deployment of ML models at scale. No-Code AI Platforms: These platforms empower users to create and deploy ML models without coding expertise. They offer intuitive interfaces and drag-and-drop functionalities, aiming to democratize AI by engaging diverse backgrounds in ML model development.

Recent AI advancements have transformed industries, yet the complexity of ML model development restricted accessibility to a specialized group. Traditional ML deployment demanded technical expertise, limiting its reach to data scientists and experts.

+ info

Objective

Enhancing Accessibility in MLOps

Task: Integrating no-code AI platforms into the realm of MLOps Goal: To significantly enhance the usability and accessibility of MLOps for individuals without extensive technical expertise. Significance: By achieving this objective, we pave the way for democratizing AI and widening its accessibility beyond the confines of technical experts.

Comparison

Implementation

Infra-as-Code & Automation

Automated Infrastructure Provisioning using IaC:

No-Code Orchestration of ML Pipelines:

• Leveraging Terraform to establish foundational infrastructure for ML pipelines.
• User-initiated configurations to create and configure Google Cloud resources.

Liberation from manual infrastructure management for practitioners.

• Integration of Terraform and Google Cloud’s Vertex AI suite.
• Initiating AutoML jobs for iterative tasks
• Programmatic deployment of generated models to endpoints.

Seamless end-to-end MLOps workflows without extensive technical specialization.

Infrastructure Setup for ML Pipelines:

1. A Google Cloud Platform project2. Enabling the required APIs 3. Service Account 4. A Cloud Storage bucket 5. A Vertex AI dataset

Manual configuration of these cloud resources can pose challenges for users unfamiliar with GCP

Model Evaluation and User Study

Predictions

User Study

Model Evaluation

Endpoint Deployment & Testing

Test ease of usability

Metrics available in Vertex AI

6 participants recruited, representing ML enthusiasts without cloud or ML expertise.Between-subjects methodology employed for three machine learning models using Terraform and Vertex AI AutoML.

• Precision
• Recall
• AuPRC
• Log loss
• Confidence Threshold
• Confusion Matrix.

These metrics provide insights into model performance and behavior.

Models deployed to endpoints enable real-time or batch predictions.Vertex AI allows synchronous or asynchronous requests for predictions.

Results

Achieved Outcomes: Usability & Satisfaction

Results showcase substantial enhancements in usability and user satisfaction validated through a user study. Statistical analysis and user feedback affirm the success of the integration.

Majority lacked prior cloud or ML experience. Participants reported satisfaction in setting up and using the MLOps pipeline. Ratings clustered around "Satisfied" and "Very Satisfied".

Statistical Analysis of SUS Score:

Descriptive Statistics:

• Median & mean SUS scores: 83.750, indicating usability.
• Moderate standard deviation suggests consistent responses.
• Normal distribution confirmed by Shapiro-Wilk test.

One Sample T-Test:

• Strong statistical significance with a t-value of 22.010.
• Very low p-value (< .001) rejects the null hypothesis.
• Participants' perceptions significantly differ from neutral usability.

Limitations exist, but future work can expand on scalability, transparency, real-time models, and customization, enhancing accessibility and democratization of AI.

Future Work

• Explore streaming data pipeline integrations.
• Integrate IaC with Cloud build
• Scale the system and optimize performance.

Short Term

• Add a web faced interface for users to limit further cloud interaction
• Personalize integrations for specific use-cases.

Medium Term

• Address ethical considerations in AI deployment.

Long Term

Goals and Aspirations

I'm eager to expand horizontally, diving into different facets of cloud and data engineering. My goal is to master various tech stacks and frameworks, creating a colorful palette to craft innovative and adaptive solutions. I aim to evolve into a mentor and guide, fostering an environment where everyone thrives on sharing knowledge and creative ideas.My ambition is to architect not just robust solutions but also sustainable ones, ensuring scalability and environmental mindfulness. I dream of leveraging technology as a force for social good, channeling my skills toward solutions that tackle real-world challenges.

Q & A

Data Collection & Analysis:

Quantitative (SUS ratings) and qualitative (thematic analysis) data collected. Quantitative data used to gauge overall system usability. Thematic analysis used to categorize qualitative responses for insights and improvements.