Arasavalli Devesh Sai Karthik

Smart BIN – AI-Powered Waste Segregation System

Smart BIN is an intelligent waste management system designed to automate waste identification and segregation, reducing the need for manual sorting. The prototype combines computer vision, embedded computing, sensor-based detection, and motorised mechanical control to create a smart and efficient waste disposal solution. The system uses a Raspberry Pi Zero 2 W with a camera module to capture images of incoming waste and classify them using a lightweight machine learning model based on MobileNetV2 and TensorFlow Lite. An Arduino Nano interfaces with sensors and actuators to coordinate the physical sorting process. Ultrasonic sensors detect incoming waste and monitor bin fullness, while servo motors control the entry and discharge mechanisms. A stepper motor rotates the platform to align the appropriate bin with the disposal funnel. The project integrates AI inference with real-time hardware control, addressing practical challenges such as motor positioning, power distribution, mechanical load, and sensor integration. Smart BIN received recognition at student innovation events, including a top-five finish at a 36-hour hackathon and selection for an Innovation Expo.

Swara OS – Telugu Voice-Driven Assistant

Swara OS is a Linux-based voice assistant project designed to enable users to interact with a computer through spoken Telugu commands. The project explores accessible, regional-language computing by combining automatic speech recognition with deterministic intent recognition and command processing. The speech-processing pipeline uses AI4Bharat's IndicConformer model through NVIDIA NeMo to convert Telugu audio into text. The resulting transcript is processed by a custom intent grammar that identifies supported command categories and extracts relevant parameters. Keyword matching and fuzzy matching using RapidFuzz help handle variations in recognised text and improve command interpretation. The project is organised into modular components for speech inference, transcript processing, and intent resolution, making it easier to test and extend the system with additional command families. Development includes Linux-based environment management, Python implementation, and Git-based version control. Swara OS aims to provide a foundation for regional-language human-computer interaction, with future scope for expanded command coverage, improved recognition accuracy, and integration with system-level actions.

AI Research Navigator – Intelligent Research Assistant

AI Research Navigator is an AI-assisted research application concept designed to simplify the process of exploring technical papers, understanding complex topics, and locating relevant information. It focuses on making research material easier to navigate through natural-language interaction and structured information retrieval. The proposed workflow uses language models to interpret user queries, identify relevant research content, and generate concise explanations based on retrieved information. A retrieval-oriented architecture can organise document content into searchable representations, retrieve relevant passages, and use those passages to support context-aware responses. LangChain and Hugging Face tools provide a foundation for integrating language models and document-processing components. The project is intended to reduce the time required to search through large amounts of technical material and help students organise their research workflow. Potential extensions include paper summarisation, topic-based discovery, source-linked answers, and comparisons of research methodologies.

Gemma – Large Language Model Fine-Tuning

This project explores the adaptation of a pretrained Gemma-family large language model for specialised tasks through supervised fine-tuning. The objective is to understand how a pretrained language model can be customised using task-specific datasets while managing GPU memory, training configuration, and model evaluation. The workflow covers dataset preparation and formatting, tokenizer configuration, model loading, training setup, and experimentation in GPU-enabled environments such as Kaggle. Hugging Face Transformers provides the model-training interface, while PyTorch supports GPU computation. Unsloth is explored to simplify and optimise fine-tuning workflows where compatible with the selected model and environment. The project also involves investigating dependency compatibility, dataset-loading issues, and training-environment configuration. These experiments provide practical exposure to the machine learning development lifecycle, from preparing training data to evaluating a fine-tuned model.

Linux-Based High-Performance Computing Cluster

This project focuses on the deployment and configuration of a Linux-based high-performance computing environment for computational workloads. It explores the system administration and networking foundations required to configure compute resources, establish connectivity, and support remote access in a multi-system environment. The work includes Linux environment setup, network configuration using Netplan, connectivity testing, and remote graphical access through X11 forwarding. It also introduces the deployment considerations involved in preparing systems for high-performance computing workloads, including network reliability, configuration consistency, and remote management. The project provides hands-on exposure to Linux infrastructure and network systems, connecting operating-system knowledge with practical deployment and troubleshooting workflows. It also supports the development of skills relevant to system engineering, infrastructure automation, and performance-oriented computing environments.

Smart Lighting and Automation System

The Smart Lighting and Automation System is an embedded electronics project focused on improving lighting control through sensor-based automation. The system is designed to reduce unnecessary manual operation by using environmental inputs to determine when lighting should be switched on or off. The proposed architecture combines a microcontroller, input sensors, and an output switching stage. Sensor readings are processed by embedded control logic, which determines the appropriate lighting state according to configured conditions. The project explores digital input/output interfacing, control logic implementation, hardware integration, and the relationship between software decisions and physical device behaviour. The design can be extended with occupancy detection, ambient-light sensing, adjustable operating thresholds, and manual override functionality. These capabilities provide a practical foundation for understanding embedded control systems and developing energy-conscious automation solutions.