Noise-Resilient AI Systems
Robust Multimodal Intelligence under Real-World Uncertainty
Live Vision Model Degradation
Interactive demonstration of robustness under visual noise.
Research Projects
A selection of my contributions towards building intelligent, reliable, and robust systems under real-world noise and uncertainty.
Robust Object Recognition Under Extreme Noise
Investigated robust limits of deep vision models (ResNet, ViT) and built Denoising Autoencoder pipelines to mitigate ~85% performance degradation under severe visual degradation.
Multimodal Memory AI: Semantic Retrieval System
Developed a memory-based AI integrating text (spaCy) and image (MobileNet) features, resolving multi-step queries and improving context handling by ~30% over stateless baselines.
Fake Image Detection System: Deepfake & Synthetic Media Analysis
Designed a detection pipeline using spatial and frequency-domain (FFT/DCT) features across 15K images, achieving ~92% accuracy in detecting GAN and diffusion-based manipulations.
Noise-Induced Hallucination in Vision-Language Models
Analyzed hallucination behavior under noisy multimodal inputs.
Memorization vs Generalization in Vision-Language Models
Explored the balance between memorization and generalization in VLMs.
Effect of Data Noise on LLM Learning
Studied the impact of noisy data on language model performance.
Personalized Memory-Aware RAG System with Context Evolution
Built adaptive retrieval systems with evolving contextual memory.
Noise-Resilient Multimodal RAG for Low-Quality Inputs
Designed robust retrieval pipelines for degraded inputs.
RAG-Driven Digital Twin for Intelligent Environments (XR + AI)
Developed a system combining XR and AI for intelligent environment simulation.
Live Commit Stream
Update ViT noise robustness benchmarks
Integrate FFT feature extraction in deepfake pipeline
Fixed context memory decay issue
Add baseline stats for Gaussian blur degradation
Refactoring RAG pipeline for Digital Twin
Technical
Expertise
Developing state-of-the-art models targeting robustness under noisy inputs and extending the boundaries of multimodal retrieval systems.
“I am interested in building intelligent systems that remain robust under noisy and imperfect real-world conditions, with a focus on multimodal learning, retrieval-augmented generation, and reliable AI.”
Ask My Research
Interact directly with my work — no need to search manually. Ask questions to receive concise, context-aware explanations strictly based on my academic portfolio.
AI Collaboration Matcher
Paste your background/project to calculate synergy score.
Awaiting Input
Let's Connect
Open to research collaborations and R&D opportunities in Multimodal AI and Robust Intelligence.
mukherjeesaptarshi289@gmail.com© 2026 Saptarshi Mukherjee. All rights reserved.
Designed for Advanced Academic R&D.