Projects

InsightForge: AI-Powered Business Intelligence Assistant with LangChain, RAG, and LLMs

Most businesses generate mountains of data — sales records, regional performance, customer demographics — but lack the tools or expertise to extract meaningful insights from it. InsightForge was built to change that: a conversational Business Intelligence Assistant that lets any user query their data in plain English and receive structured, actionable insights in seconds. Built on LangChain and LangGraph, InsightForge combines Retrieval-Augmented Generation (RAG) with OpenAI GPT-4 to ground every response in real business data rather than hallucinated answers. A three-way query router intercepts each request before any LLM call — directing chart requests to a ChartEngine, aggregation queries to DuckDB SQL, and open-ended reasoning to the full RAG pipeline. This dramatically reduces latency and API costs while maintaining response quality. The system also features a dual-layer memory architecture for coherent multi-turn conversations, semantic chunking by business domain for precise retrieval, and a QAEvalChain evaluator that scores every response across accuracy, completeness, relevance, and business value. The frontend is built with Next.js 13 and Nivo Charts, backed by a FastAPI REST API and PostgreSQL with pgvector for semantic search. Click here for detailed post.

CustomerSupport AI: Multi-Agent Customer Support System with Claude, LangGraph, RAG & MCP

When a bank customer fires off "My debit card hasn’t arrived in three weeks!", something deceptively complex needs to happen — the system must classify the complaint, empathise, generate a unique ticket, persist it to a database, and respond warmly, all in under four seconds with no human in the loop. Customer Support AI is a production-grade multi-agent customer support system built as an Applied Generative AI capstone that does exactly this. The system is a directed graph of eight specialised AI agents orchestrated by LangGraph, each owning exactly one job — classification, feedback, RAG policy Q&A, ticket lookup, ticket closure, and fallback escalation. Every routing decision is an explicit Python conditional, not an LLM inference, making the system fully auditable — critical for a regulated banking domain. A FAISS-backed RAG pipeline with confidence gating prevents policy hallucination, while a Model Context Protocol (MCP) tool layer centralises all business logic — ownership validation, SLA breach detection, and PII masking — in a single place every agent shares. This post walks through every architectural decision in detail: why LangGraph over CrewAI or AutoGen, why MCP over direct DB access, why RAG with a confidence gate over prompt-stuffing, and the six complete data flows that handle every customer scenario from positive feedback to ownership-validated ticket closure. Click here for detailed post.

NewsGenie: Building a Real-Time AI News Chatbot with Multi-Agent Architecture

Most AI chatbots today can generate fluent answers, but they struggle with one critical problem — staying up to date. Ask about current events, and you risk getting confident but outdated or even fabricated responses. This is where NewsGenie was born: an AI-powered news chatbot designed to combine the reasoning capabilities of large language models with real-time information retrieval. Instead of relying on a single monolithic model, NewsGenie takes a different approach — a coordinated team of specialised agents, each responsible for a specific task such as intent classification, news retrieval, validation, and response composition. Orchestrated using LangGraph, this architecture brings structure, reliability, and scalability to AI systems that need to reason and act in real time. The system integrates multiple data sources like the GNews API for live headlines and DuckDuckGo as a fallback, ensuring graceful degradation when external services fail. It also incorporates session memory for conversational continuity, a two-pass classification system for accurate intent detection, and credibility validation for every news article surfaced to the user. In this post, I walk through the complete technical design of NewsGenie — from high-level architecture and agent workflows to real-time data integration and error handling. It’s a deep dive into how modern AI systems can move beyond simple text generation and evolve into structured, reliable, and production-ready applications. Click here for detailed post.

5-Day AI Agents Intensive Course with Google

Artificial Intelligence is evolving from systems that simply generate text to ones that can reason, plan, act, and learn. This shift marks the rise of AI agents — intelligent, goal-oriented systems that bridge the gap between models and autonomous behavior. Google’s 5-Day AI Agents Intensive, hosted on Kaggle, captured this transformation in a hands-on, engineering-focused program that went far beyond traditional AI courses. Over the course of five days, the workshop unpacked the full lifecycle of agentic systems — from foundational reasoning loops and tool integration to long-term memory, evaluation, and large-scale deployment. Powered by Gemini and the Agent Development Kit (ADK), it showed how to design, observe, and productionize AI agents that don’t just respond but actually think and act. What made this program exceptional was its focus on real-world implementation. Participants explored topics like the Model Context Protocol (MCP) for interoperability, AgentOps for observability and reliability, and Vertex AI Agent Engine for production deployment. Each concept was tied to code labs, whitepapers, and evaluation exercises that mirrored Google’s own internal frameworks. This article is a summary of my journey through the course — what I built, what I learned, and how it reshaped my understanding of building trustworthy, scalable AI systems. Click here for detailed post.

Flowise No Code Multi-agent LLM Interview Question Generator

The Interview Question Generator Agents project is an AI-driven multi-agent system that generates Java programming interview questions, test cases, and solutions. Flowise, a no-code AI workflow builder, orchestrates LLM-powered agents for seamless task execution. The Supervisor Agent dynamically assigns tasks to worker agents, ensuring efficient question generation, validation, and solution refinement. Flowise enables LLM orchestration, API integration, memory management, and task automation with no-code drag-and-drop functionality. It optimizes AI workflows, reducing complexity while improving accuracy. With agent-based execution and structured decision-making, Flowise enhances scalability, automation, and efficiency. Click here for detailed post.

Multi-agent LLM automation system built to assist in automating academic and career-related tasks

AI Student Helper is a multi-agent automation framework utilizing LangGraph to orchestrate intelligent task execution for academic and professional assistance. It employs a Supervisor Agent powered by LLM-based intent recognition to route user requests dynamically to Specialized Agents, each leveraging NLP, OCR, web automation, and API integrations for task execution. The PDF Summarizer Agent extracts and processes text using PyMuPDF and LLM-based summarization models, while the YouTube Summarizer Agent utilizes speech-to-text conversion via Whisper AI for transcript generation. The Job Application Agent automates profile matching and form submission using Selenium and LinkedIn API, while the Homework Helper Agent provides interactive AI-driven problem-solving via Streamlit. The graph-based execution model ensures optimized task prioritization and resource-efficient processing, integrating AI-powered decision trees and modular workflow automation for seamless user experience. Click here for detailed post.

AWS Rekognition and Java to Identify Objects / Text in Images

Amazon Rekognition is an Amazon Web Service (AWS) that provides image and video analysis services. You can provide an image or video and the service will detect objects, people and scenes. The purpose of Rekognition is to analyze images and predict what objects are in the image, if there are any faces, transcribe text, and other tasks.In this post, we can see the demo of AWS Rekognition which using both custom model & predefined models. Click here for detailed post.

Deep Learning for NLP: Twitter Sentimental Analysis in Python

Obtaining feedback from a few of the customers is easy. However, imagine that you could ask the whole world! You would know immediately the general opinion and could react accordingly. This machine learning solution examines twitter data for their tonality. The classification, whether a tweet is positive or negative, i.e. the tonality, is done by a deep neural network. The deep convolutional neural network architecture we constructed in this project,is a feedforward neural net using the TensorFlow library in Python. The neural network classifies tweet vectors in negative or positive using the Bag of Words model,ReLU activation function and softmax activation function, which will classify the tweet to either positive or negative.Click here for code and detailed report.

Real Time Stock Market Prediction - Neural Network, ARIMA, RandomForest, SVM, DecisionTree

According to the forecast of stock price trends, investors trade stocks. In recent years, many researchers focus on adopting machine learning (ML) algorithms to predict stock price trends. In this work, Artificial Neural Network /ARIMA / Random Forest / SVM/ Decision Tree techniques have been utilized for predicting the next day closing price for companies in real time.To achieve this goal, we trained traditional machine learning algorithms and created/trained multiple deep learning models taking into consideration the importance of the stock indicators & historical data. Click here for detailed post.

R - Apriori - Online Buying Behavior

Ability to recognize and track frequent item sets in data help businesses shift through the layers of seemingly unrelated data for meaningful relationships. Apriori algorithm is one of the most popular data mining approach for finding frequent item sets from a transaction dataset and derive association rules. Rules are the discovered knowledge from the data base. Once frequent item sets are obtained,it is straightforward to generate association rules with confidence larger than or equal to a user specified minimum confidence.In this post we can explore how to find association rules and formulate some predictions for the online store. Click here for detailed post.

R - Neural Network & SVM - Predict Success Of Bank Telemarketing

The shift towards decision making (DM) based on machine learning techniques techniques will change the way in which consumer markets and our societies function. Through ML, predictive analytics is being used by businesses to identify these patterns and major trends with the objective to improve the DM and influence future business outcomes. This project proposes an Neural Network (ANN)/ SVM approach to predict the success of telemarketing calls for selling bank long-term deposits. To validate the proposed model, we uses the bank marketing data of 41188 phone calls. The NN attains 90% of accuracy which outperforms other conventional classifiers and confirms that it is credible and valuable approach for telemarketing campaign managers. Click here for code and detailed report.

R - Naive Bayes - Validating Yelp Food Establishment Ratings based on Reviews.

We are living in a review-obsessed culture and, when we are looking for a place have lunch or have dinner, it has now become a common reflex to check online ratings before making a final call.It is obvious that a restaurant with poor rating will probably face serious business issues and maybe have to stop its activity. By building Rating Prediction which will relate text based feedback to a star rating, we can help these restaurants understand how they would rate on Yelp .Our approach is to use Naive Bayes classifier and build a text classifier which predicts the number of stars given by the user based on the review text. Click here for code and detailed report.

R - KNN - Predicting Song Popularity using Spotify Dataset

Why Do Some Songs Become Popular? DJ Khaled boldly claimed to always know when a song will be a hit. We decided to further investigate by asking three key questions: Are there certain characteristics for hit songs, what are the largest influencers on a song’s success, and can old songs even predict the popularity of new songs? Predicting how popular a song will be no easy task. To answer these questions, we made use of the Spotify song dataset, and used KNN machine learning to build a model with 85% accuracy. Click here for code and detailed report.

Analysing Cryptocurrency Market- Linear, Polynominal, Spline, GAM, Cubic Regression & Extreme Gradient Boosting

Cryptocurrencies are becoming increasingly relevant in the financial world and can be considered as an emerging market.By analyzing bit coin historical features, such as bitcoin tradevolume, bitcoin blockssize, bitcoin difficultyto find a new block,total value of coinbase block rewards , transaction fees paid to miners, we can predict the variation and can predict the ups and downsof bitcoin price . In this post, we propose the usage of common machine learning algorithms like Linear, Polynominal, Spline, GAM, Cubic Regression & Extreme Gradient Boosting for predicting the price movement of the Bitcoin. Click here for detailed post.

Python - ML - Predicting video game hits

Hundreds of new games launch every month. Yet only a fraction of makes their way to the top. If you ever wondered what makes a video game a bestseller, now you have a full overview of factors that make its breakout success. In this project, I will analyze sales data from 8K video games, identify variables most correlated to hits (games which sell over 1M units) and implement a prediction model using machine learning algorithms to separate winners from losers. Click here for code and detailed report.