Decoder-Only Transformer Implementation (Group Project)
Participated in the development of a decoder-only Transformer model, implementing key components of modern sequence modeling. Contributed to the design and implementation of:
Multi-Head Self-Attention: Captures dependencies across input sequences for contextual representation.
Feed-Forward Layers: Applies nonlinear transformations to enhance feature extraction.
Residual Connections and Layer Normalization: Ensures stable training and efficient gradient flow in deep architectures.

I worked as part of a team on the backend development of Easy-Rent, a no-broker house rental application designed to simplify the rental process for users. Collaborating closely with my teammates, I contributed to building a reliable and scalable backend using Node.js with the Express framework.
My responsibilities included implementing authentication and authorization mechanisms, integrating payment services, enabling map-based location features, and developing automatic email notifications. I also implemented rate limiting, robust error handling, and efficient search functionality to ensure system security, performance, and a smooth user experience. This project strengthened my experience in collaborative development and in building production-ready backend systems.

Fine-Tuned LLaMA + RAG for an AI-Driven E-Learning Platform (iCog Labs)
During my work at iCog Labs, I contributed to building an LLM-based intelligent assistant for an e-learning platform, designed to help learners interact with course content through natural language.
Problem:
Learners needed personalized, context-aware explanations and content retrieval from large volumes of educational material, which static search systems could not provide effectively.
Solution & Techniques Used:
Fine-tuned a LLaMA model on domain-specific educational data to improve response relevance and pedagogical quality.
Implemented RAG using LangChain, allowing the model to retrieve relevant learning materials before generating answers.
Designed prompt strategies focused on clarity, step-by-step reasoning, and learner-friendly explanations.
Evaluated responses for accuracy, relevance, and hallucination reduction.

I worked this project for Beta Tech Hub, Here are the list of tasks that I have done:
● Architecting a scalable, multi-tenant IDS/IPS system using Kubernetes and
Kubebuilder. Developed custom controllers to automate the deployment of
isolated, per-tenant security instances, ensuring strict data segregation and
resource efficiency.
● Integrating machine learning–based detection and prediction models to
enhance the accuracy and intelligence of security tools.
● Designing and implementing microservice architectures to integrate multiple
open-source security tools, enabling unified data collection and optimized
threat analysis.

Intrusion Detection and Prevention System using Deep
Learning
● Designed and developed an end-to-end, machine learning–driven IDS/IPS to
detect and prevent cybersecurity threats.
● Built the system from scratch, implementing packet capture and network
interface control in C, and feature extraction, model inference, process
management, and user interface in Python.
● Developed a custom flow-based feature extractor, adapted from
CICFlowMeter, to support accurate attack classification.
● Trained a Deep Neural Network using TensorFlow on the UNSW-NB15 dataset to
detect multiple attack categories, including anomalous traffic patterns.
● Implemented inter-process communication between C and Python
components using the ctypes library for efficient data exchange.
● Developed a desktop monitoring interface using Tkinter for real-time
visualization and system control.
● Applied knowledge from programming, machine learning and cybersecurity,