Posts by Collection

portfolio

Customer Churn Prediction Analysis Using Ensemble Techniques in Machine Learning

Project Code: Visit
Customers are a company’s most valuable asset, and maintaining customers is critical for any organisation looking to increase revenue and develop long-term meaningful relationships with customers. Furthermore, the cost of obtaining a new client is five times that of keeping an existing customer. Customer Churn/Attrition is one of the most well-known business difficulties in which consumers or subscribers discontinue doing business with a service or a firm. Ideally, they will no longer be a paying customer. A client is considered to have been churned if a certain length of time has passed since the consumer last interacted with the company. Identifying whether or not a client will churn and offering relevant information aimed at customer retention are crucial to lowering churn. Our brains cannot anticipate customer turnover for millions of clients; here is where machine learning may assist.

Amrita Canteen App

Project Code: Visit
Web App. built using advanced DL techniques to provide insights into the menu & crowd statistics of the canteen using CCTV cameras. Under review of the college board, to be adopted in canteens.

Hyperparameter Optimization with Genetic Algorithm and Simulated Annealing

Get pip package: Visit
This repository Link contains a Python package jss-optimizer for optimizing hyperparameters using genetic algorithm (GA) and simulated annealing (SA) hybrid optimization algorithm.

Sea Dragon 🐉 - Autonomous Underwater Vehicle

The Sea Dragon is an advanced Autonomous Underwater Vehicle (AUV) designed for underwater exploration and data collection. Built using state-of-the-art components, including an ESP32, Raspberry Pi 4B, and a variety of sensors, the Sea Dragon is capable of executing complex missions in challenging underwater environments.

FotoFind: Revolutionizing Image Retrieval with Advanced Computer Vision and NLP Techniques

The dramatic rise in digital image creation has presented modern systems with massive repositories of visual data. Managing, indexing, and retrieving relevant images is crucial now due to the various forms in which photos are stored, from personal photo libraries to organizational databases, scientific imagery surveillance data, etc. Conventional solutions often rely on user-generated tags or rudimentary filenames, providing only limited search capabilities. Such manual metadata tagging can be inconsistent and quickly becomes infeasible for extensive datasets. To address these bottlenecks, FotoFind implements a pipeline of advanced computer vision and natural language processing models to glean semantic information from images automatically. The synergy of object detection, captioning, and OCR ensures that each image, regardless of its content, gains enriched metadata that can be leveraged to facilitate more nuanced text-based searching. The aim is to streamline and automate a process that would otherwise depend heavily on manual labor.

Unlocking Efficiency in Implicit CoT Reasoning: A Speculative Decoding Approach

This project introduces a novel speculative decoding technique tailored to enhance the efficiency of chain-of-thought (CoT) reasoning in large language models (LLMs). CoT reasoning enables models to solve complex tasks through sequential reasoning steps, but it is often hindered by high memory demands and inter-token latency during top-k decoding. Our method leverages inherent parallelism in CoT tasks, enabling simultaneous token generation while maintaining high reasoning accuracy. By combining open-source models like Qwen2.5 and Vicuna 7b and conducting experiments on datasets such as GSM8K and StrategyQA, we demonstrate that our approach significantly reduces computational overhead without compromising performance. This advancement opens new pathways for deploying LLMs in real-world applications requiring efficient and accurate reasoning.

publications

AI based parameter estimation of ML model using Hybrid of Genetic Algorithm and Simulated Annealing

Published in 14th ICCCNT conference, Indian Intitute of Technology Delhi, India, 2023

In this study, parameter estimation of ML models, here specifically, random forest classifier was conducted on a heart disease dataset. We have suggested a new approach to machine learning that hybridizes genetic algorithm with simulated annealing for estimating the hyperparameters of the random forest classifier. This application of hybrid optimization helped in increasing the accuracy of the machine learning model by 10% and is very promising in comparison to various different classification methods for this same problem.

Download here

Enhancing AUV Sensor Precision with Adaptive Genetic Algorithm aided Kalman Filtering

Published in 15th ICCCNT conference, Indian Intitute of Technology Mandi, India, 2024

Accurate sensor data is crucial for the successful operation of Autonomous Underwater Vehicles (AUVs). However, dynamic underwater environments and inherent sensor limitations pose significant challenges. This paper investigates using the Kalman filter to enhance instantanoius sensor data accuracy in AUV navigation by integrating with Genetic Algorithms (GAs). The proposed Adaptive Genetic Algorithm-based Kalman Filter (AGAKF) adjusts filter parameters in real-time using GAs as compared to the existing methods that use batch processing. Extensive simulations and comparisons show that AGAKF achieves superior noise reduction and better signal preservation than the existing techniques, enhancing AUV navigation accuracy in diverse underwater environments.

Download here

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015