NISHU KUMAR

AI Engineer

[email protected] +91 7549453235

Resume

Hello, I am

NISHU KUMAR.

3

GenAI
Projects

1

Deep Learning
Project

2

Machine Learning
Projects

About Me

IIT Delhi & IIT Jodhpur graduate and AI/ML Engineer specializing in production-grade agentic systems and RAG pipelines using LangChain,
LangGraph, Google ADK and Agno. Expertise in multi-agent architectures, LLM fine-tuning, and scalable backend deployment using FastAPI,
Docker, and AWS.

Key Skills

Programming Languages: Python, C/C++, SQL, PowerShell

Frameworks and Platforms: Agno, LangGraph, LangChain, FastAPI, Streamlit, Anaconda

AI/ML Expertise: Agentic AI Systems, Multi-Agent Architectures, LLM Fine-tuning, RAG Pipelines (ChromaDB), Knowledge Graphs, NLP

Data and ML/DL Tools: Pandas, NumPy, TensorFlow, PyTorch, Neural Networks

Databases and Cloud: MySQL, SQL Server, SQLite, Vector Databases (ChromaDB), Docker, AWS EC2

My Projects

E-commerce Chatbot

Domain/Function: Retrieval-Augmented Generation

Multi-Agent Data Science Team

Domain/Function: Multi-Agent System

Research Assistant

Domain/Function: Retrieval-Augmented Generation

AI-Powered Vehicle Damage Detection System

Domain/Function: Computer Vision

Credit Risk Classification

Domain/Function: Finance/Credit Risk Modelling

Health Insurance Cost Predictor

Domain/Function: Health Care/ML

My Experience

Augment Systems — AI Engineer

Gurgaon | Sep 2025 – Present

• Designed and deployed an end-to-end Generative AI chatbot system using LLMs and Retrieval-Augmented Generation (RAG) to deliver accurate, con responses across diverse user queries.

• Architected a production-ready conversational AI pipeline using Llama 3.3 via Groq API, implementing intelligent query classification that achieved 90%+ intent detection accuracy.

• Built a RAG pipeline with ChromaDB vector database enabling semantic search over FAQ documentation and product knowledge to generate grounded, hallucination-resistant responses.

• Developed a natural language to SQL engine that converts user questions into executable SQL queries, enabling real-time product search across 1,000+ inventory items stored in SQLite.

• Implemented a semantic query routing system using Sentence Transformers (all-MiniLM-L6-v2) to automatically direct queries to appropriate pipelines (FAQ retrieval, SQL generation, or conversational responses).

• Built an interactive LLM-powered web application using Streamlit, supporting streaming responses, session management, and responsive UI for real-time user interaction.

• Implemented secure database access controls restricting operations to read-only SELECT queries, preventing SQL injection and unauthorized data manipulation.

• Developed a data ingestion and processing pipeline by scraping e-commerce product data from Flipkart, performing cleaning, deduplication, and migration into SQLite for efficient querying.

• Optimized vector embeddings and semantic retrieval algorithms, achieving sub-second response latency for knowledge retrieval operations.

Impact: Delivered a scalable AI-powered customer query automation system improving response accuracy, reliability, and system security.

Blu Parrot Technologies — GenAI Intern

Remote | June 2025 – Aug 2025

• Built an AI-powered HRMS platform using Claude AI and FastMCP, automating 15 HR workflows including employee onboarding, leave management, meeting scheduling, and internal ticketing.

• Developed multi-step backend orchestration workflows integrating HRMS systems, SMTP notifications, equipment provisioning, and calendar scheduling with conflict detection.

• Implemented natural language search capabilities using fuzzy matching and hierarchical manager relationships for intelligent employee data retrieval.

• Designed a secure credential distribution system using Gmail SMTP with TLS encryption, ensuring safe onboarding communication and account setup.

Impact: Reduced employee onboarding processing time by ~95% through AI-powered workflow automation.

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

Natural Language Processing

Gen AI to Agentic AI with Business Projects

The_Ultimate_Job_Ready_Data_Science_Course_Certificate

AI Agents using Agno

Docker for Machine Learning

Let's Connect

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.

Mail Me

[email protected]

https://www.linkedin.com/in/nishu73/

Call Me

+91 7549453235

GitHub

https://github.com/NISHU8875

GenAI Project

E-commerce Chatbot

Domain/Function: Retrieval-Augmented Generation

Developed a production-ready e-commerce chatbot using Llama 3.3 via Groq API, integrating RAG with ChromaDB, LLM-to-SQL for real-time product search across 1,000+ items, semantic routing, and a secure, interactive Streamlit interface with low-latency responses.

View Project
on GitHUB

View Project
on Streamlit

Other Tools

Generative AI

Large Language Models (LLM)

Vector Databases

SQLite

Streamlit

Python

Project Landing Page: Streamlit interface for "ShopAssist," a GenAI chatbot. Features a user-friendly UI for seamless interaction, intent-based routing, and real-time query handling using Llama 3.3 and Groq API for optimized performance.

Live Product Query Execution: Demonstration of Text-to-SQL capabilities. The bot processes natural language for "Nike shoes under budget," executes a secure SQLite query, and retrieves real-time data with direct Flipkart links and product ratings.

Project Details

🤖E-commerce Chatbot (GenAI & RAG)
Developed an intelligent E-commerce assistant that utilizes Retrieval-Augmented Generation (RAG) and Agentic workflows to automate customer support and product discovery.

✅ Project Highlights:

• Designed a semantic routing system using Hugging Face embeddings to classify queries into FAQ (RAG), Text-to-SQL product search, and conversational flows.

• Implemented a ChromaDB-based RAG pipeline for grounded, hallucination-free responses from company knowledge.

• Engineered a Natural Language to SQL engine enabling real-time inventory search across 1,000+ products using SQLite.

• Integrated Llama 3.3 via Groq API for low-latency inference and streaming responses.

• Developed a responsive Streamlit UI and automated product data ingestion via web scraping.

GenAI Project

Multi-Agent Data Science Team

Domain/Function: Multi-Agent System

Engineered a hierarchical multi-agent system using Agno and Llama 3.3 to automate end-to-end data science workflows. Features a lead "Manager Agent" orchestrating 6 specialized sub-agents for automated EDA, ML pipeline generation, and real-time Streamlit data visualization.

View Project
on GitHUB

Other Tools

Multi-agent Systems

Autonomous Agents

Agno

Data Science

Python

Agent Orchestration: Backend view of a hierarchical multi-agent system on Agno. Shows the Manager Agent delegating tasks to specialized sub-agents (Data Loader, Viz Agent, Coding Agent) to autonomously process a 40,000+ row dataset.

Automated Insights: Autonomous Exploratory Data Analysis (EDA) execution. The multi-agent system processes raw CSV data and programmatically generates complex Matplotlib visualizations and statistical reports through natural language prompts.

Project Details

🤖Multi-Agent Data Science System
Developed a sophisticated Multi-Agent Data Science System designed to transform natural language prompts into production-ready ML pipelines.

✅ Project Highlights:
• Hierarchical Orchestration: Architected a "Team Leader" agent to decompose complex user requests and delegate tasks to 6 specialized agents, including Data Analysts and Viz Agents.
• Agentic Communication: Implemented a centralized state-management system using SQLite, allowing agents to persist memory and share dataframes across different stages of the workflow.
• Domain-Specific Toolsets: Built custom tool integrations for Pandas, Scikit-learn, and Matplotlib, enabling autonomous exploratory data analysis (EDA) and model training.
• Modular Infrastructure: Leveraged the Agno framework to ensure a model-agnostic backend, supporting Llama 3.3 (via Groq) and OpenAI.

GenAI Project

Research Assistant

Domain/Function: Retrieval-Augmented Generation

An AI-powered RAG system that responds to the user query just by adding the URL. The system fetches the webpage, chunks the text, and stores it in Chroma DB. Extract similar chunks for the query and pass them to LLM to generate source-grounded output.

View Project
on GitHUB

View Project
on Streamlit

Other Tools

Lang Chain

Hugging Face Embedding

Streamlit

Python

Chroma DB

Initially, it processes the URLs first by accessing the data ingestion pipeline. Once the data is ingested, it provides us with the relevant answers.

Here's the next question, asked by the bot, about the model capabilities and the purpose of launching the model.

Project Details

Manually scanning articles, Wikipedia pages, and long-form blogs is slow and inefficient. This project addresses that gap by building a Research Assistant powered by Retrieval-Augmented Generation (RAG), enabling users to submit up to three URLs and receive precise answers grounded strictly in those sources.

Features:
• Drop and process the URL to fetch article content.
• The content is fetched using Web Base Loader powered by Langchain.
• The extracted content is divided into chunks using the Recursive text splitter.
• Construct a vector store to store embeddings of these chunks into ChromaDB to enable swift retrieval of relevant content.
• Queries are created using the Langchain Prompt template and provide flexibility to the user and their queries.
• Interacts with the LLM(llama 3 via Groq) by inputting queries to receive answers with the source URL.

Deep Learning Project

AI-Powered Vehicle Damage Detection System

Domain/Function: Computer Vision

Developed an AI-based system using CNN and transfer learning to detect and classify damage into six categories. The model helps automate vehicle inspection with an accuracy of 80% and is deployed using Streamlit.

View Project
on GitHUB

View Project
on Streamlit

Other Tools

Convolutional Neural Networks (CNN)

Computer Vision & Deep Learning

Transfer Learning

PyTorch

ResNet50

Model Deployment (MLOps)

Streamlit

The model shows this car has "Rear Crushed" damage.

The model shows this car has "Front Breakage" damage.

Project Details

This Project, "Car Damage Detection", aims to automate the process of visual inspection of car damage and classify the damage using Deep learning. This model is trained on a 2300-labelled car image dataset. To improve model performance, extensive data augmentation was performed, including contrast adjustment, rotation, and colour jittering.

Model Details:
1. It can classify the damage into six categories:
•Front Normal
•Front Breakage
•Front Crushed
•Rear Normal
•Rear Breakage
•Rear Crushed
2. The model uses a transfer learning approach with pre-trained weights of ResNet50.

Tools & Tech Stack:
• Torch & Torch Vision Training CNN models, Image transformations and data loading.
•Matplotlib: Visualizing different car images.
• EfficientNet-B0, ResNet50: Using Pre-trained weights with the help of transfer learning.
• Optuna: Hyper Parameter Tuning
• Scikit-learn: Model evaluation through accuracy and confusion matrix.

Machine Learning Project

Credit Risk Classification

Domain/Function: Finance/Credit Risk Modelling

Built a Streamlit app that automates the process of credit risk evaluation using machine learning. This app is designed for a loan processing officer to check an applicant's creditworthiness for the applied loan by providing the default probability, credit score and rating associated with it.

View Project
on GitHUB

View Project
on Streamlit

Other Tools

Machine Learning

Exploratory Data Analysis

optuna

streamlit

Significantly low Loan-to-income ratio resulting in less default probability.

Increased Loan-to-income ratio and higher delinquent months are more prone to becoming a defaulter

Project Details

This app enables a finance company to quickly assess the credit risk of loan applicants, reducing decision-making time while improving accuracy. It predicts the probability of default, generates a credit score (300–400) with corresponding ratings, and provides clear insights for decision-making.
The major contributors to finding the default probabilities are: Loan-to-income ratio, Credit utilization ratio and Delinquency ratio.

Features:
• Probability of the applicant being a defaulter
• Generates credit score & rating for easy interpretation
• Achieved >90% accuracy and 93% recall for identifying defaulters
• Strong model performance with AUC 98% and KS Statistic 85.9
• Built with Logistic Regression, XGBoost, SMOTE Tomek, Optuna

Tools & Tech Stack:
• Python(Programming Language)
• Numpy, Pandas, Matplotlib, Seaborn (Data Visualization & Analysis)
• Sickit-learn, XGBoost, Optuna (Model training and tuning)
• Ml-flow (Experiment Tracking)
• Streamlit (Deployment)

Machine Learning Project

Health Insurance Cost Predictor

Domain/Function: Health Care/ML

This Streamlit web application predicts the expected health insurance premium based on a user’s personal, health, and lifestyle details. It helps users estimate their annual premium by leveraging a machine learning model trained on real-world data.

View Project
on GitHUB

View Project
on Streamlit

Other Tools

Machine Learning

Exploratory Data Analysis

XGBoost

streamlit

Python

A patient is having a lower age, no medical history does not have much impact on the premium, but bmi category of Overweight increases some premium.

In this case a patient has relatively more age, with a severe medical history, and a gold plan increases the premium amount.

Project Details

⚕️Health Insurance Cost Predictor:
Manually determining health insurance premiums is slow and inconsistent. To address this, I built an end-to-end ML application that automates premium prediction for each patient, significantly reducing decision-making time while maintaining high accuracy and low error.

✔️ Key Features:
•Risk Score Generation: Computes a medical risk score based on disease history and patient attributes.
• Age-Segmented Models: Uses two optimized models for different age groups (>18 and <25) to improve reliability.
• High Performance: Achieves 97% accuracy with an error rate under 10%.

🔍 Model Highlights:
Engineered a normalized risk scoring system to quantify cumulative medical impacts, optimizing premium forecasting accuracy. By integrating genetic risk factors for the under-25 demographic, the model refined actuarial precision and reduced outlier prediction error to 2.5%.