Mohamed Zayaan S

AI Researcher at IIT Madras specializing in Trustworthy AI, Mechanistic Interpretability, and Embodied Intelligence.

My goal is to build the next generation of safe and beneficial AI systems.

About Me

I am a pre-final year B.Tech Civil Engineering student at IIT Madras, with a minor in Computing/CS. My journey into AI is driven by a fascination with the fundamental principles that should guide its creation. I don't just want to build capable systems; I want to build systems that are robust, fair, and provably beneficial to humanity.

Experience

I have been fortunate to contribute to cutting-edge research and product development across academia and startups, building a comprehensive understanding of AI systems.

Center for Responsible AI (CeRAI), IIT Madras

Research Intern | July '25 - Present

Architecting novel bias classifiers to detect and mitigate implicit socio-cultural biases within LLMs, leading to an upcoming publication.

Advanced Geometric Computing Lab, IIT Madras

Research Intern | April '25 - Present

Implementing skeletal-prior-embedded attention models for creating robust 1D curve skeletons from 3D point-cloud registration and developing transformer-based geometric feature learning.

FinMitr (Startup)

Co-founder & AI Head | June '24 - Feb '25

Led the AI/ML team, architecting and building FinGuru, an AI-powered financial advisor, and worked on business operations strategy.

SPIRE Lab, IISc Bangalore

Research Intern | Nov '24 - Jan '25

Engineered an advanced audio preprocessing pipeline that reduced model complexity by 40% for machine learning applications in signal analysis.

Key Projects

A selection of projects where I have built complex, novel systems from first principles.

Circuit Vision: Mechanistic Interpretability

Reverse-engineered sentiment pathways in LSTMs, achieving a 71.2% isolation quality for the target circuit. This work moves beyond correlation to a causal understanding of a model's internal reasoning.

PyTorch LSTM Interpretability

NeuroGenesis: Self-Evolving SNN

Architected a self-evolving Spiking Neural Network that dynamically adapted its own structure. Achieved >94% accuracy gain over baseline in a continual learning setting while reducing memory footprint by 60MB.

Spiking Neural Networks Continual Learning Computational Efficiency

Swarm Robotics Coordination

Developed a novel multi-agent hybrid-RL algorithm for swarm communication by fusing Reinforcement Learning with Ant Colony Optimization. This innovation now forms the basis of a filed patent.

Patent Filed Reinforcement Learning Swarm Intelligence

AI-driven Video Captioning

Implemented a novel 2-stage fusion architecture using BLIP for visual analysis and Flan-T5 to refine captions with Whisper audio context, drastically reducing factual hallucination.

Multimodal AI Transformers BLIP Flan-T5

Quantum-AI Optimizer

Engineered a quantum-classical optimizer for Max-Cut/TSP problems using QAOA and a DQN agent for adaptive variational tuning, cutting runtime by 70% and converging 3x faster than baselines.

Quantum Computing QAOA DQN Qiskit

Search Retrieval Engine

Engineered an information retrieval pipeline from scratch using Latent Semantic Analysis (LSA) via SVD to overcome Vector Space Model limitations, achieving statistically significant retrieval gains.

Information Retrieval LSA SVD

Publications & Patent

Upcoming Publication

A Novel Bias Classifier for Mitigating Socio-Cultural Biases in LLMs

Work from my research at the Center for Responsible AI (CeRAI), IIT Madras.

Patent

Hybrid TD3-ACO based Swarm Robotic system for Intelligent Warehouse Automation

Status: Submitted for review.

Get In Touch

I'm always open to discussing new research, projects, or opportunities.