Software Engineer

Naheed Rayan

Passionate about building intelligent systems that solve real-world problems. I bridge the gap between cutting-edge AI research and scalable software engineering, with expertise in LLM/RAG systems and published research in machine learning.

View Projects
Naheed Rayan

Path

Experience

AI enginner

2024 — 2025

AI Engineer at SimplySolve Inc.

Developed document RAG systems on AWS Bedrock for automated analytics. Built CSV analytics pipeline with custom LLM integration for data visualization.

Backend Engineer

2025 — Now

Shikho

Leading backend development and AI/ML R&D for Shikho's educational platform. Built comprehensive RAG pipeline for Shikho AI, executed large-scale data migration from legacy systems, and optimized codebase performance for enhanced user experience.

Showcase

Selected Works

Shikho AI RAG Pipeline
test1test2

Shikho AI RAG Pipeline

Built end-to-end RAG (Retrieval-Augmented Generation) pipeline for Shikho's AI-powered educational platform. Led data migration from legacy Arango database to MongoDB and implemented comprehensive codebase optimizations.

Knowledge

Research

2025

Transient Turn Injection: Exposing Stateless Multi-Turn Vulnerabilities in Large Language Models

Introduces Transient Turn Injection (TTI), a novel multi-turn attack technique that systematically exploits stateless moderation by distributing adversarial intent across isolated interactions in LLMs. Key Contributions: Developed automated black-box evaluation framework for testing LLM adversarial robustness Evaluated across state-of-the-art models from OpenAI, Anthropic, Google, Meta, and open-source alternatives Uncovered significant variations in resilience and previously unknown model-specific vulnerabilities Proposed practical mitigation strategies including session-level context aggregation

View Paper
2023

Latent Representation and Generative Augmentation with Graph-Based Learning for Imbalanced Leukocyte Cytomorphology Classification

Proposed a deep learning framework combining convolutional autoencoder embeddings, GAN-based minority class augmentation, and GraphSAGE classification. Achieved 91.6% accuracy and improved macro-F1 on the AML-Cytomorphology-LMU dataset, with enhanced recognition of rare cell types and more stable cross-validation performance.

View Paper

Journal

Latest Articles

Test Title
February 2026

Test Title

Inquiry

Let's talk.

Available for new projects, full-time opportunities, or just a coffee chat.

Get in Touch