Portfolio - Lalith D

Hello, I am

Lalith D

Data Professional

About Me.

Data professional with solid experience working across the data lifecycle — from collecting and processing data to analyzing it and turning it into meaningful insights. I enjoy solving problems with data, building reliable workflows, and continuously learning new tools and techniques along the way.

2

Machine Learning Projects

3

GenAI Projects

1

Deep Learning Project

Skills.

Programming : Python, SQL

Machine Learning & AI : Scikit-learn, XGBoost, TensorFlow, PyTorch, Model Training, Evaluation, Optimization

Generative AI & NLP : LLMs, Prompt Engineering, RAG Pipelines, Semantic Search, Embeddings, CrewAI

Computer Vision : YOLO, Image Classification, Object Detection, Transfer Learning

MLOps & Deployment : CI/CD Pipelines, Model Monitoring, Docker, MLflow

Big Data : Apache Spark , Databricks, Hadoop

Cloud Platforms : AWS, Azure OpenAI

Frameworks & Tools : LangChain, LangGraph, CrewAI, FastAPI, Flask

Projects.

My Experience.

Awards & Certificate.

Python: Beginner to Advanced For Data Professionals

SQL for Data Science

Math and Statistics For AI, Data Science

Master Machine Learning for Data Science & AI: Beginner to Advanced

Deep Learning: Beginner to Advanced

Introduction to IOT - CISCO

Programming in JAVA - NPTEL (IIT Kharagpur)

AI Agent Fundamentals - Databricks

Generative AI Fundamentals - Databricks

Contact.

Feel free to get in touch with me. I am always open to discussing new projects, creative ideas or opportunities to be part of your visions.

[email protected]

https://www.linkedin.com/in/lalithkumard/

GitHub

https://github.com/lalithdabilpuram01

GenAI Project

Agentic Youtube RAG Pipeline

Domain/Function: Agentic AI

An autonomous multi-agent AI pipeline built with CrewAI and RAG. It orchestrates intelligent agents to dynamically ingest YouTube channel transcripts, perform contextual research, and synthesize technical blog articles without human intervention.

View Project
on GitHUB

Other Tools

CrewAI

Python

RAG

Showcasing a multi-agent workflow using CrewAI to automate content creation. The terminal displays a Blog Researcher agent extracting insights from YouTube videos, which a Blog Writer agent then transforms into a structured technical post.

Demonstrates the CrewAI Blog Writer agent autonomously scraping YouTube channel metadata. It captures the agent’s ability to search specific video sources, retrieve relevant content, and summarize complex ML topics into a professional final output.

Project Details

Architected and developed an autonomous, multi-agent AI pipeline using the CrewAI framework and Retrieval-Augmented Generation (RAG). This system automatically ingests, analyzes, and synthesizes YouTube channel content to generate comprehensive, highly polished technical blog articles without human intervention. I engineered a sophisticated orchestration system consisting of specialized AI personas (like a Senior Content Researcher and a Technical Writer) that delegate analytical tasks sequentially. By integrating advanced RAG methodologies to dynamically scrape large-scale video transcripts from YouTube, the AI retrieves hyper-relevant insights based on natural language queries. I leveraged state-of-the-art LLMs (GPT-4o ecosystem) to ensure high-fidelity comprehension of complex engineering topics. This project successfully demonstrates advanced proficiency in generative AI workflows, agentic automation, and transforming unstructured multimedia data into structured narratives.

GenAI Project

News Research Tool

Domain/Function: Generative AI / Natural Language Processing (NLP)

Built a cost-efficient GenAI news research tool using a RAG pipeline. Processes live URLs, filters relevant content via embeddings and FAISS, and delivers source-grounded answers while reducing token usage and hallucinations.

View Project
Engagement

View Project
on GitHUB

Other Tools

LangChain

FAISS

OpenAI

API

dotenv

Pickle

Streamlit

Dashboard visualization for a News Research Tool built with LangChain and OpenAI. Highlights a Retrieval-Augmented Generation (RAG) architecture used to transform complex articles into a searchable, accurate, and cost-efficient knowledge base.

User interface for an AI-powered news analysis tool. It features a sidebar for URL management and a main query engine that uses Retrieval-Augmented Generation to answer complex questions based on real-time web articles.

Project Details

Built a Retrieval-Augmented Generation (RAG) application to analyze live news articles from multiple URLs.

Implemented intelligent text chunking to split long articles into semantically meaningful segments.

Generated vector embeddings and stored them in a FAISS vector database for efficient semantic retrieval.

Retrieved only the most relevant content chunks per query, reducing token usage and inference costs by up to 90%.

Designed a retrieval-constrained LLM pipeline to produce accurate, source-grounded answers with citations.

Added persistent vector storage to avoid redundant processing and improve performance across sessions.

Developed an interactive Streamlit interface for end-to-end user interaction and research workflows.

Machine Learning Project

Credit Risk Modeling

Domain/Function: Finance / Banking / Risk Analytics

Built a machine learning model using XGBoost and Optuna to predict loan default risk with 92% AUC accuracy, enhancing credit decisioning and risk transparency through SHAP explainability and a Streamlit web app.

View Project
Engagement

View Project
on GitHUB

View Project
on Streamlit Cloud – Interactive web app for real-time credit risk prediction.

Interactive Streamlit app for credit risk prediction. Input loan details, get instant default probability, credit score, and risk rating.

Streamlit app highlighting a high-risk credit scenario with specific inputs. Shows 99.98% default probability, Credit Score 300, and a "Poor" rating,

Project Details

Conducted Exploratory Data Analysis (EDA) using Pandas and NumPy to understand borrower demographics, income levels, and loan performance patterns.

Performed extensive data preprocessing — handling missing values, encoding categorical features, and scaling numerical variables to ensure model stability.

Built multiple classification models including Logistic Regression, Random Forest, and XGBoost, with Optuna used for hyperparameter tuning to maximize accuracy.

Evaluated model performance using metrics such as AUC-ROC, precision, recall, and F1-score, achieving a strong AUC of 0.92 on test data.

Applied SHAP interpretability to explain model predictions, providing transparency in credit decisioning.

Deployed the model using Streamlit for an interactive user interface where users can input loan attributes and instantly view predicted risk categories.

Deep Learning Project

Vehicle Damage Detection

Domain/Function: Computer Vision

Built a real-time car damage classifier using ResNet-50 and Transfer Learning. Classifies 6 damage types from vehicle images. Deployed as a live web app on Streamlit Cloud. Tech: Python, PyTorch, Streamlit.

View Project
on GitHUB

View Project
on Streamlit Cloud

Other Tools

Pytorch

Pandas

Optuna

Matplotlib

Streamlit

Home page of the Vehicle Damage Detection app. Users can upload a vehicle image (JPEG, PNG, JPG) via drag and drop or file browser. The model then analyzes the image and returns an instant damage classification.

The model in action. A heavily damaged car image is uploaded and the system instantly predicts the damage as "Front Breakage" — identifying structural and mechanical damage at the front of the vehicle with high accuracy.

Project Details

Built a real-time car damage classification system using ResNet-50 and Transfer Learning on ImageNet weights. The model is fine-tuned to classify vehicle images into 6 categories — Front Normal, Front Breakage, Front Crushed, Rear Normal, Rear Breakage, and Rear Crushed. Implemented a complete inference pipeline using PyTorch and Torchvision covering image preprocessing, normalization, and prediction. Deployed as a live web application on Streamlit Cloud for instant damage assessment. Designed with a production-first mindset, targeting real-world use cases in vehicle rental return workflows and insurance claim automation.

Machine Learning Project

Health Insurance Premium Prediction

Domain/Function: Healthcare / Insurance Analytics

Developed a machine learning model using XGBoost to predict health insurance premiums based on personal and lifestyle factors. Improved model accuracy by 28% through feature engineering and cross-validation, and deployed it on Streamlit Cloud for real-time premium estimation.

View Project
Engagement

View Project
on GitHUB

View Project
on Streamlit Cloud

Streamlit app for health insurance premium prediction. Input demographic and health factors (age, gender, smoking) to get accurate cost estimates.

Health Insurance Cost Predictor in action! After inputs, the Streamlit app shows a predicted premium of 4004, demonstrating real-time cost estimation.

Project Details

Objective:
Developed a machine learning model to predict individual health insurance premiums based on demographic and health-related factors, helping insurers and users estimate fair pricing.

Approach:

Performed data cleaning and EDA to explore relationships between age, BMI, region, and smoking habits.

Built regression models including Linear Regression, Random Forest, and XGBoost to predict premium costs.

Used feature engineering and cross-validation to improve model accuracy and reduce overfitting.

Achieved a 28% reduction in Mean Squared Error (MSE) through hyperparameter optimization.

Deployed the model on Streamlit Cloud, creating an interactive app where users can input personal attributes to get real-time premium predictions.

Outcome:
Delivered a reliable, interpretable, and user-friendly predictive system that helps users understand how various factors influence insurance costs.

GenAI Project

AI Philosophy Chatbot

Domain/Function: Generative AI / NLP

Built a Streamlit web app integrating OpenAI GPT-3.5 to answer user questions from the Bhagavad Gita. Features a custom-styled interface and real-time spiritual responses.

View Project
Engagement

View Project
on GitHUB

A Streamlit web app where users ask questions and receive AI-generated answers from the Bhagavad Gita, set on a divine-themed interface.

AI chatbot app answering user questions from the Bhagavad Gita in real time, showing thoughtful responses on a themed Streamlit interface.

Project Details

Objective:
To build an interactive web application that provides answers inspired by the Bhagavad Gita using an AI-powered conversational interface.

Approach:

Integrated OpenAI’s GPT-3.5 model to generate thoughtful and verse-based responses to user questions.

Built the frontend with Streamlit, including a visually styled layout featuring dynamic background images, transparent cards, and responsive input/output sections.

Implemented user input handling and real-time answer rendering through the ask_gita() function, which interacts directly with the OpenAI API key defined in openai_helper.py.

Applied structured prompts that ensure responses align with the teachings of the Bhagavad Gita while maintaining modern language clarity.

Outcome:
Developed a user-friendly chatbot capable of answering spiritual questions in real time with context drawn from Bhagavad Gita principles.