Portfolio - Sharanch Mukhia

GenAI Project

Stateful-Chatbot

Domain/Function: Conversational AI System Engineering

A production-ready conversational chatbot built with LangGraph featuring real-time streaming, persistent SQLite-backed memory, resumable chat threads, tool calling, and full observability via LangSmith — all integrated into a clean, interactive Streamlit UI.

View Project
on GitHUB

View Project
on

Other Tools

LangGraph

Langchain

LangSmith

SQLlite

Streamlit

GROQ

Python

Interface

Workflow

Project Details

I wanted to go beyond simple chatbot demos and build something that feels like a real product. Many chat apps reset when refreshed, forget past conversations, and don’t show what’s happening behind the scenes. My goal was to create a chatbot that responds in real time, remembers conversations, lets users continue previous chats, uses helpful tools like search, and still feels clean and easy to use.
I designed the system to handle conversations smoothly, save chat history automatically, and allow users to resume any past session. I added live streaming so replies feel faster and more natural, and built in clear tracking so it’s easy to understand and improve how the system works. The result is a reliable, user-friendly chat application that feels polished, responsive, and ready for real-world use.

GenAI Project

TubeRAG

Domain/Function: Generative AI

Developed a Retrieval-Augmented Generation (RAG) system that processes YouTube video transcripts, stores them in a vector database, and enables context-aware question answering using large language models. The system retrieves the most relevant transcript chunks via semantic search and grounds LLM r

View Project
on GitHUB

View Project
on

Other Tools

Youbtube Transcript API

Streamlit

Output

Architecture

Project Details

Extracting specific information from long-form YouTube content such as podcasts, lectures, and tutorials is time-consuming and inefficient when done manually. The goal of this project was to build an automated system that enables fast, accurate, and reliable question answering directly from video content while minimizing large language model (LLM) hallucinations.

To address this, I developed a Retrieval-Augmented Generation (RAG) system using LangChain, YouTube Transcript API, ChromaDB, and GPT-4o-mini, incorporating MMR-based semantic retrieval with optimized chunking (500 characters with 80 overlap) and deploying the solution through an interactive Streamlit web interface. The system reduced information retrieval time from 15–20 minutes to under 10 seconds per query, achieved 85–90% accuracy on factual questions, processed videos up to 3+ hours, and lowered hallucination rates from 30–35% to under 4% across 100+ evaluated question–answer pairs.

Machine Learning Project

Credit Risk Modeling - Classification

Domain/Function: Financial Services (NBFC)

Designed a credit risk solution for an NBFC to identify high-risk borrowers early in the loan lifecycle. The model enables risk-based approvals and pricing, reduces potential default losses, and strengthens overall asset quality.

View Project
Engagement

View Project
on GitHUB

Other Tools

Optuna

Streamlit

A credit risk assessment interface where user inputs borrower, loan, and credit behavior details to calculate default risk. The tool supports real-time, risk-based loan approval and pricing decisions.

A credit risk modeling interface that captures borrower demographics, loan characteristics, and credit behavior variables to estimate probability of default, derive a credit score, and assign a risk rating for underwriting decisions.

Project Details

Lauki Finance Company, a fictional NBFC, aimed to strengthen asset quality by reducing loan defaults and improving risk-based lending decisions. I built an interpretable, end-to-end credit risk model using customer, loan, and bureau data, with domain-driven feature engineering such as loan-to-income and delinquency ratios. Class imbalance was handled using Borderline SMOTE, and performance was optimized with Optuna. The final model achieved an ROC-AUC of 0.98, Gini of 0.96, and a KS of 85.7%, with the top two deciles capturing 98% of defaulters, enabling risk-based approvals and pricing while reducing credit losses and strengthening asset quality.

Machine Learning Project

Premium Prediction - Regression

Domain/Function: Healthcare

Built an end-to-end machine learning solution for insurance premium prediction, focusing on data quality, feature engineering, model evaluation, and error diagnostics to ensure reliable, business-aligned pricing beyond aggregate performance metrics.

View Project
Engagement

View Project
on GitHUB

Other Tools

Streamlit

This is a Health Insurance Cost Predictor that estimates a user’s insurance premium based on demographic, financial, and health details. Users enter their information, and a model predicts and displays the expected insurance cost.

Workflow

Project Details

Insurance companies require accurate premium pricing to balance competitiveness and profitability, yet manual pricing is slow and inconsistent. Even with automation, high aggregate metrics can be misleading. In this project, a unified model achieved an R² of 0.97, but deeper evaluation revealed it was not reliable for real-world pricing.
Residual and percentage-error analysis uncovered a 33% extreme mispricing rate (>10%), concentrated among young customers (≤25). I implemented an age-based segmented modeling strategy, building a dedicated young-customer model with medical and genetic risk features and validating it using cross-validation and error diagnostics. This reduced extreme errors from 33% to ~3%, eliminated severe mispricing, and delivered a production-ready, business-aligned pricing system with lower revenue risk.

Python Project

Expense Tracking System

Domain/Function: Software Development

A full-stack expense tracking application that records daily expenses, categorizes spending, and visualizes spending patterns through an intuitive and interactive web interface.

View Project
Engagement

View Project
on GitHUB

Other Tools

MySQL

FastAPI

Streamlit

Users can add and update expenses for a selected date, as well as retrieve all expense records associated with that specific date.

Users can select a start and end date to view category-wise expense breakdowns, shown through interactive charts and a clear summary table.

Project Details

Faced with the challenge of inefficient manual expense tracking, I set out to build a system that simplifies logging and analyzing daily spending. I developed a full-stack Expense Tracking System using FastAPI, MySQL, Streamlit, RESTful APIs, and intuitive CRUD features for adding, updating, and retrieving expenses by date. I also built interactive visualizations and summary tables that provide category-wise spending insights for any selected date range. As a result, users can effortlessly track expenses, understand their financial patterns, and make more informed budgeting decisions, resulting in ~90% faster expense monitoring and clearer spending insights.

Hello, I am

Sharanch Mukhia

1

2

2

About

Key

My

Stateful-Chatbot

TubeRAG

Credit Risk Modeling - Classification

Premium Prediction - Regression

Expense Tracking System

Contact Me

Email

Github

Linkedin

Call

Sharanch Mukhia

Download Resume